The submodules code caches data about submodules in a way that
can cause problems.  This adds some tests that try making various
modifications to the state of a submodule to see where we can
catch out problems in the submodule caching.

Right now, I've put in an extra git_submodule_reload_all so that
the test will pass, but with that commented out, the test fails.
I'm working on fixing the broken version of the test at which
point I'll commit the fix and delete the extra reload that makes
the test pass.