GTID 5/5: GTID-aware replica catchup

[driv3r]

Stacked PR 5 of 5 — GTID support

Base: #448. Final PR in the stack.

What this PR does

Makes WaitUntilReplicaIsCaughtUpToMaster coordinate-mode aware so running Ghostferry from a replica works in GTID mode, using executed-set containment instead of file/position ordering.

• Adds ReplicatedMasterCoordinateFetcher + ReplicatedMasterGTIDViaCustomQuery, and IsCaughtUpToCoordinate (GTID containment / file-pos compare). The legacy file/position fetcher and IsCaughtUp API are preserved via an adapter.
• Ferry passes BinlogCoordinateMode to the wait object; the init probe uses a mode-appropriate coordinate.
• copydb and sharding gain ReplicatedMasterGTIDQuery and select the GTID fetcher in GTID mode.
• Tests: GTID replica-catchup suite; replication-config and master-change tests made mode-aware.

Verified locally on MySQL 8.0: full Go and Ruby suites pass in both file_position and gtid modes.

---

Whole-stack validation summary

• test/go: pass in both modes; copydb/test + sharding/test pass.
• Ruby suite: file_position 73 tests/650 assertions, gtid 73 tests/651 assertions — 0 failures.

[driv3r]

Review follow-ups addressed (stack-wide)

Aggregated the three reviews and pushed fixes to the relevant stage branches (rebased through the stack). Summary:

#447 (gtid-stage3-streamer)

• Empty-set stop hang (High): shouldContinueStreaming no longer treats an empty stop coordinate as "unset". FlushAndStop always records the target before stopRequested, so an empty executed set on a fresh source is a valid, already-reached target instead of an infinite wait.
• Non-XID GTID advancement (High): DDL/admin QueryEvents (which commit without an XIDEvent) now advance the streamed GTID set; BEGIN/COMMIT/ROLLBACK are skipped. Prevents cutover hangs when the trailing statement is DDL.
• gtid_mode validation (Medium): CheckServerGTIDModeEnabled is now wired into Ferry.Initialize for source and (when verifying) target.

#446 (gtid-stage2-coordinate-mode)

• Reparse on hot path (Medium): added NewGTIDCoordinateFromSet, which caches the parsed mysql.GTIDSet; HasReached reuses it, so the per-event stop check no longer serializes+reparses. Cache is unexported / not serialized.

#448 (gtid-stage4-state-resume)

• Resume mode mismatch (High): Ferry.Start rejects resuming when the state's coordinate mode differs from the configured mode.
• Empty-set target-verifier resume (High): presence is checked via HasTargetVerifierBinlogCoordinate() (pointer/field), not IsZero(), so an empty GTID coordinate isn't mistaken for absent.
• Fail-closed intersection (Medium): intersectGTIDSets/MinSourceBinlogCoordinate return errors on parse/set-math failure instead of silently falling back to one side; Ferry.Start propagates.
• Interface breakage (Low/Med): SetCoordinates removed from the exported DMLEvent interface; stamping uses an internal coordinateStamper assertion.
• CI (Medium): go-test now runs gtid on MySQL 8.0/8.4 (file_position on all).

#449 (gtid-stage5-replica-progress)

• Empty ReplicatedMasterGTIDQuery (and copydb's position query) rejected at init.
• Master @@GLOBAL.gtid_mode=ON validated in GTID mode.
• Docs: BinlogCoordinateMode comment updated for shipped behavior; documents that legacy file/position progress fields are not populated in GTID mode (use coordinate fields).

Added tests: empty stop target, QueryEvent advancement, transaction-control detection, NewGTIDCoordinateFromSet clone/cache + JSON exclusion, multi-UUID intersection, fail-closed parse error, target-verifier presence with empty set.

Validation (MySQL 8.0.32): full Go test/go + copydb + sharding pass in both modes; full Ruby suite passes in both modes (file_position 73/650, gtid 73/651). Build + gofmt clean.