Support namespaced references again

[ethomson]

Support git_repository_set_namespace again (the equivalent of GIT_NAMESPACE); this was subtly broken with worktree support.

This adds a simple namespace test (equivalent to rugged's) and simplifies the namespace handling by simply appending the namespace (if it exists) to the gitdir and commondir.

/cc @carlosmn

[ethomson]

So there are two things here that probably bear discussion:

1. The interaction between namespaces and worktrees. If you set a namespace, this will now look for namespaces in both. It's not exactly clear that this is what git itself would do, because git's namespaces are very narrowly applied to "server" functionality. I don't know that there are commands that obey namespaces and worktrees both. So I think our implementation is okay, but if there exists a command in the future (or in the present, and I just don't know it) that obeys both then we may need to rethink this.

2. You can create a refdb that is not backed by a filesystem path. (See refs::iterator::empty, for example.) We handle this (and have since the inception of the refdb) by simply creating paths like refs/heads/master (instead of prefixing them with the gitdir). This is obviously pretty wrong. I haven't fixed this here, since it's a) a longstanding bug not a regression and b) I want to unblock Update libgit2 to b31b236... rugged#687. But I will open a PR to deal with this when I have a few spare minutes.

[carlosmn]

Namespaces are such an underused feature that I'm happy with just proving the minimal support works. We can worry about their interaction with worktrees if someone ever actually cares.