erlc_compile, for before/after compiling .erls to
.beams, but before .app.src to .app

this ensures the project apps are compiled to `lib/` instead of `extras/`

Several projects use an include path relative
to the project's root.
file:compile will look in three places for the include
files:
The current working directory
The directory where the module is being compiled
The directories given by the include option

this ONLY attempts to deduplicate test sets that are generated by
rebar in the absence of any user specified tests

If a bad configuration file is submitted to rebar3 shell, display the
following error:

    ===> The configuration file submitted could not be read and will be
    ignored.

And keep going otherwise rather than silently failing.

While crash-fast is usually a good mechanism, the shell so far is very
tolerant of failures from apps to boot and whatnot, so this feels
appropriate.

Fixes #1019

`beam_lib:chunks(..)` needs a path to object code which, frustratingly,
`code:which/1` won't return for cover compiled modules. instead just
assume that if `code:which/1` doesn't return `non_existing` a module
is something we can run tests on

this is slower than the compile time check but i guess packaging rebars with
repos is still a thing and i think only the eunit and ct providers call it
anyways

into directory test arguments

- robocopying a directory into another directory recursively expects the
  directory name to be properly mapped onto the destination, otherwise
  all the files are copied into the given path. This patches things so
  a directory-to-directory robocopy works as expected in a linux mindset
  so tests pass
- the test for canonical paths didn't expect a windows environment at
  all; the test (and library) is modified to be consistent in that
  environment: always with a native format and with proper support of
  drive letters.

This reuses the trick used within OTP to pick within old and new shell.

The 'user' structure is the same for all cases (escript, escript + dumb
TERM, unstable install, unstable install + dumb TERM), so we take it
down first.

Then we boot the TTY driver, which fails if TERM=dumb, in which case we
boot the retro-style usr. If it worked, we shut down the driver again,
and boot a modern shell structure.

This avoids all warnings and seems to work in all cases.

This is quite the hack.

This requires to detect the current shell running; if it's the new
shell, business as usual.

However, if it's the old shell, we have to find a way to take over it
and drive IO. This requires a few steps because:

- the old shell does not let you be supervised intelligently (it uses
  supervisor_bridge, so killing the child is not a supported operation
  from the supervisor)
- the old shell ignores all trappable exit signals except those coming
  from the Port in charge of stdio ({fd, 0, 1})
- the old shell shuts down on all exit signals from the stdio Port
  except for badsig, and replicates the shutdown reason otherwise
- An escript does not tolerate the `user` process dying (old shell) for
  any non-normal reason without also taking the whole escript down
- Booting in an escript has an implicit 'noshell' argument interpreted
  by the old shell as a way to boot the stdio Port with only stdout
  taken care of

Because of all these points, we have to kill the old `user` process by
sending it a message pretending to be the Stdio port dying of reason
`normal`, which lets it die without triggering the ire of its
supervision tree and keeping the escript alive. This, in turn, kills the
old stdio port since its parent (user.erl) has died.

Then we have to boot our copy of user.erl (rebar_user.erl) which
conveniently ignores the possibility of running the stdio port on stdout
only -- always using stdin *and* stdout, giving us a bona fide old-style
shell.

A known issue introduced is that running r3:do(ct) seems to then kill
the shell, and r3:do(dialyzer) appears to have an odd failure, but
otherwise most other commands appear to work fine.

if these directories actually exist they'll be added to the path ahead
of the release/standard distribution directories and they'll break eunit
and/or ct execution

fixes #950

this allows repeated test suite names across apps without conflicts

previously rebar3 dropped suites declared at the root of the project (via
`--suite=whatever_SUITE' probably) and warned. this was because the compiler
would recursively copy and compile everything in the directory indicated by
the test suite. this changes the copy mechanism to only copy erl source files
and directories that end with `_SUITE_data' into the `extras' dir in `_build'

This lets a plugin define templates to be loaded:

    $ rebar3 new
    ...
    proper (plugin): A basic PropEr suite for an OTP application
    ...
    $ rebar3 new help proper
    proper:
            plugin template (...)
            Description: A basic PropEr suite for an OTP application
            Variables:
                    name="suite" (...)
                    ...
    → rebar3 new proper fakesuite
    ===> Writing test/prop_fakesuite.erl

In this case, proper is a fake template file I've put by hand in
_build/default/plugins/rebar3_proper/priv/<somename>/, meaning it will
only work as far as it's called from the project's root.

The priority order of plugins is now .config > plugin > built-in, such
that someone could ensure plugins do not crush their own private
templates, but also that custom or plugin templates do overtake built-in
ones. It used to be Built-in > .config only.

Templates are searched for recursively in the priv/ directory of
plugins.

This should fix #955

The test is implicit as a bad index previously silently crashed rebar3.
By adding the bad index to the `new` suite's files, we can show that
things keep running.

this restores compatibility with rebar2 and erlang.mk

Full rewrite, code should be understandable now.

- proper segregation of comparison between tuple terms and non-tuple
  terms. Guards weren't specific enough and that meant the wrong clauses
  of guards would be triggered
- proper deduplication of entries in the list. An additional N passes
  are required (we co-opt the reverse step to be more efficient) because
  while the original lists:umerge easily removes dupes, this is
  requiring more logic here since `[a,{a,b},{a,b,c},a,{a,b,c}]` is a
  possible interleaving and we'd want `[a,{a,b},{a,b,c}]` -- comparison
  of direct neighbours isn't enough.

The previous default meant that verbose output would not be emitted.

When compiling a dependency with a MIB file the generated hrl file is left in
the root project directory in a file called "include". This has the perverse
effect of messing up the search path for include files causing any dependencies
with files in their "include" directory to fail to build after that.

Relx config need new values at the end of the list and
profile additions prepended. So now in apply_profile in
the case it is a relx config we prepend the config list.
This does, however, mean that a profile release can not
'extend' a release defined in the default profile since
the one being extended would have its definition after
the profile release and thus relx would not be able to
use its values for extending.

The checking of flags and the parsing of arguments is separated up.