import: a migrated merge request has something to diff !243

Stacked on !242. Last of the v1.13.0 code.

The bundle carried no head, so every replayed merge request had an empty one. mr diff resolves through refs/merge-requests/<n>/head, never set, and a merged MR has no source branch to fall back on — so it could only fail. The bundle now carries head_sha, merged_base and merged_at, a merged MR is replayed with MarkMerged (not SetMRState, which leaves the base empty), and the ref is pointed at the head once objects exist.

Re-running an import repairs the ref: that path used to continue on an already-imported MR before reaching any of this, which meant the bundle-then-push order — the normal one — could never produce a working diff.

GitHub half: the head was already recorded, and the ref set when objects happened to be present, which for a mirror made with default refspecs they are not. The import fetches refs/pull/*/head into refs/gh-pull/* first, one fetch rather than one per PR, token via GIT_ASKPASS like the mirror worker rather than in a URL in argv. It reports how many MRs ended with no head objects.

Verification, honestly. You said to attempt both halves knowing the GitHub path can't be exercised from here. The migrate half is covered by TestMigratedMRHasADiff against two real instances — and worth noting, my first version of that test passed without the fix, because with the source branch still present the diff resolves through it and an empty head is invisible. It uses a merged MR with a deleted branch now and fails without the fix; I checked. The GitHub path is written against the documented refs/pull/* behaviour and is unverified until you run a real import.

Closes #128