Since the lager_format rewrite, ~p and ~w now have different behaviour
with regard to printing lists as strings.

If the result of the string format was longer than the remaining length,
the returned remaining length would be negative, which would make very
strange things happen.

Before this fix, the code assumed that processes without a registered
name omitted the {registered_name, X} tuple in the crash report. In
actual fact, process_info/2 returns [] when the process has no
registered name, so unregistered processes were being printed with a
"name" of "[]". This corrects the code to print the Pid in this case
instead.

Specifically, the handling of lists/binaries/atoms is closer to
io_lib:format's behaviour, and some regular expressions are eliminated.

In string mode, "..", '..' and <<"..">> are stripped, just like with
io:format.

As Ryan pointed out, the 'name' of a supervisor's child may be
'undefined' for a simple_one_for_one supervisor.

The ~w formatting used before was not printing lists embedded in other
terms. This change uses trunc_io to print those terms safely and without
potentially introducing linebreaks (like ~p would)

The test if an element in a list was in the printable range didn't
handle the case of a float in the list that compared to be in the
printable range.

This is superior to float_to_list as it only prints the meaningful
digits and 1.0 isn't rendered as "1.00000000000000000000e+00".

Use lager_trunc_io:format() for all messages submitted via
lager:log().  Also, refactor list flattening and final newline
stripping into shared func that uses re:replace() to do both.

If backends have a failure or something, don't discard the messages,
wait half a second and then try sending them again.

This solves an issue reported by 'cloudhead' in IRC where deep lists
weren't being flattened when printed with format specifiers like ~s ~p
and ~P, all of which should flatten the list.