Introduce ref cursors for repository parsing

Summary:
Ref T4327. I want to make change parsing testable; one thing which is blocking this is that the Git discovery process is still part of PullLocal daemon instead of being part of DiscoveryEngine. The unit test stuff which I want to use for change parsing relies on DiscoveryEngine to discover repositories during unit tests.

The major reason git discovery isn't part of DiscoveryEngine is that it relies on the messy "autoclose" logic, which we never implemented for Mercurial. Generally, I don't like how autoclose was implemented: it's complicated and gross and too hard to figure out and extend.

Instead, I want to do something more similar to what we do for pushes, which is cleaner overall. Basically this means remembering the old branch heads from the last time we parsed a repository, and figuring out what's new by comparing the old and new branch heads. This should give us several advantages:

• It should be simpler to understand than the autoclose stuff, which is pretty mind-numbing, at least for me.
• It will let us satisfy branch and tag queries cheaply (from the database) instead of having to go to the repository. We could also satisfy some ref-resolve queries from the database.
• It should be easier to extend to Mercurial.

This implements the basics -- pretty much a table to store the cursors, which we update only for Git for now.

Test Plan:

• Ran migration.
• Ran bin/repository discover X --trace --verbose on various repositories with branches and tags, before and after modifying pushes.
• Pushed commits to a git repo.
• Looked at database tables.

Reviewers: btrahan

Reviewed By: btrahan

CC: aran

Maniphest Tasks: T4327

Differential Revision: https://secure.phabricator.com/D7982