Don't over-aggressively clean the code path in the presence of lib_dir directives

[Vagabond]

Rebar, when it encounters a lib_dir directive, caches the current code
path, adds the libdir(s) and returns the cached copy of the path. When
rebar has finished processing that directory, it restores the cached
path. This is problematic in the below scenario:

        /(lib_dir)->G
  A -> B -> C -> D -> E
   \-> F -> D -> E

When rebar is finished processing B, it restores the code path to what
it was before it processed B, removing C, D, E and G from the code path.
This means when it comes to process F, neither D or E are in the code
path, so any header includes, rebar plugins or parse transforms will not
be in the code path. Without the lib_dir directive, rebar does no code
path cleanups, so everything works fine.

This change makes rebar only remove the explicit lib_dir code paths it
added and adds an inttest that replicates the above scenario.

I don't actually know why rebar feels the need to do this in the first place, but this code has been in rebar since 2009, pretty much unchanged:

3990f0a

This patch still ensures the lib_dirs are stripped when exiting the directory, but if the lib_dirs contain applications with deps those will not be removed from the code path. I don't know if this is good or bad.

[Vagabond]

cc @jaredmorrow

[Vagabond]

Actually, to clarify, D will be added to the code path, but since D is now in the list of 'processed' directories, it will be 'skipped' and its deps won't be evaluated, so E will not be re-added to the code path.

[jaredmorrow]

We have been running this code for a month in basho/rebar and has been tested aggressively.