Running tests on windows creates false positive due to setup failures
not executing all tests.  Checking for tmp_dir was failing on
subsequent runs.  Added cases to tmp_dir creation to delete preexisting
tmp_dirs in .eunit directory.

This change fixes #310.

Modules that include the EUnit header get an implicit test/0 fun,
which cover considers a runnable line, but eunit:(TestRepresentation)
never calls.  Result: prod modules with tests can never reach 100%
coverage.  Ironic.  In this case, fix it by decrementing the NotCovered
counter returned by cover:analyze/3.

1.  When running the eunit command with the convention of putting
tests in "*_tests" modules, eunit would run those tests twice.  This
is because: 1) eunit:test/1 will naturally look for foo's tests both
in foo, and in foo_tests, and 2) eunit:test/1 was being folded over
all project modules.  The fix is to filter "*_tests" modules from the
list passed to eunit:test/1.

2.  When running the eunit command with cover enabled and tests in a
'test' directory, cover would error because it couldn't find the
source code for those tests.  This is because cover:analyze/3 will
only find module source in "." and "../src".  This is hard-coded in
cover :-(.  Since cover shouldn't be calculating code coverage on test
code anyway, the fix is to not fold cover:analyze/3 over
non-production code.

3.  When running the eunit command with cover enabled and a test suite
defined, cover would only attempt to calculate coverage on the the
test suite itself.  This was because only the suite was passed to
cover:analyze/3.  The fix is to fold cover:analyze/3 over all the
production code, filtering out the suite module if it is defined.