if used as context managers, the parsers will automatically release their file locks.

We keep stdout closed, which seems to have the side-effect of
stdout being connected to your TTY, in case you run a terminal.

However, this shold also prevent deadlocks, as only stderr is used.
The alternative would have been to try to fetch lines concurrently,
and we have been there.

For clone(), `communicate()` is used, and with some luck this will
just do the right thing. Even though last time I checked, it
didn't ... ? Lets see.

Stab at #72

Similar to git, we now ignore options which have no value.
Previously it would not handle it consistently, and throw a parsing
error the first time the cache was built.

Afterwards, it was fully usable though.
Now we specifically check for the case of no-value options instead.

Closes #349

When the code is run without setting up loggers, the loggers have no
handlers for the emitted messages. The logging module displays:
`No handlers could be found for logger "git.cmd"` on the
console. By adding a NullHandler (a no-op) the message disappears,
and doesn't affect logging when other handlers are configured.

* config parser now handles quoted values correctly. This doesn't hamper
  multi-line support.
* added regression test to travis to assure we will be warned if we
  rewrite and break the user's .gitconfig file
* only rewrite configuration files if we actually called a mutating
  method on the writer. Previously it would always rewrite it.

Fixes #285

Previously, the implementation would gladly reset new commits in submodules,
and/or reset a dirty working tree.

Now the new force_reset/force flag has to be specified explicitly to get back
to the old behaviour.

All submodule tests except for one are working.

Assured unicode values are supported when reading the configuration,
and when getting author/committer information from the environment.

Fixes #237