This changes the get_entry() method to return a refcounted version of
the config entry, which you have to free when you're done.

This allows us to avoid freeing the memory in which the entry is stored
on a refresh, which may happen at any time for a live config.

For this reason, get_string() has been forbidden on live configs and a
new function get_string_buf() has been added, which stores the string in
a git_buf which the user then owns.

The functions which parse the string value takea advantage of the
borrowing to parse safely and then release the entry.

It fails at least on my computer, though it may depend on some unpredictable
factors (say, will realloc() extend the memory segment in place, or it will
allocate new memory).

* Error-handling is cleaned up to only let a file-not-found error
  through, not other sorts of errors.  And when a file-not-found
  error happens, we clean up the error.
* Test now checks that file-not-found introduces no error.  And
  other minor cleanups.

For example, if you have

[include]
path = foo

and foo didn't exist, git_config_open_ondisk() would just give up
on the rest of the file.  Now it ignores the unresolved include
without error and continues reading the rest of the file.

We have been refreshing on read and write for a while now, so
git_config_refresh() is at best a no-op, and might just end up wasting
cycles.

There are a number of tests that modify the global or system
search paths during the tests.  This adds a helper function to
make it easier to restore those paths and makes sure that they
are getting restored in a manner that preserves test isolation.