Write commit hashes into the lock dir when locking git repositories

[Leonidas-from-XIV]

This PR applies #13176 and adds the remaining code to implement the behavior discussed in #13110. The tests change accordingly with the changed semantics of git URLs.

Closes #13110

[Alizter]

This case happens when it is a dune pin or it is a files directory opam pin. Is that what we want here?

[Alizter]

I can't work out how this can be None, if it is really impossible we should Code_error instead of falling back.

[Leonidas-from-XIV]

I've added a Code_error.raise but the tests in opam-source-conversion.t are failing. Need to investigate this, because they might be resolving the packages incorrectly as well.

[Leonidas-from-XIV]

Ok, so this branch triggers when this is happening in opam-source-conversion.t:

  $ mkpkg testpkg <<EOF
  > url {
  >   src: "git+http://github.com/foo"
  >   checksum: "md5=069aa55d40e548280f92af693f6c625a"
  > }
  > EOF

This is because the URL is a Git URL, but the resolved package testpkg actually comes from the file system (the mock repo), thus it is loaded via Resolved_package.local_fs which does not load an At_rev.t (and should not need to).

It seems like relying on the Resolved_package to read the revision from is wrong in the general case, as the Resolved_package only matches the source when pinning or vendoring but does not when the package is loaded from an OPAM repo (at which point the Resolved_package is in the OPAM repo and completely disconnected from the actual source).

[Alizter]

Can you add test cases for branches and tags?