feat(mongodb-reactive): add audit compaction capability

[davidfrigolet]

Summary

• Implements AuditCompactor for the MongoDB reactive audit store, mirroring the MongoDB sync implementation (fix(mongodb-sync): audit compaction capability #967).
• No rekey needed: current-state and ledger forms share the same auto _id, filtered by changeId, so compaction only deletes superseded documents.
• Survivor is addressed by its stored _id, not by attribute matching.
• Bridges the reactive driver's Publisher-based find()/deleteMany() to the synchronous compact() contract via the existing PublisherSync idiom.

Test plan

• Unit tests against a mocked reactive MongoCollection
• MongoDBReactiveAuditCompactionConformanceTest against a real MongoDB container, plus a store-specific assertion that the survivor keeps its original _id
• All compaction test classes (SQL, Couchbase, DynamoDB, MongoDB sync, MongoDB reactive) run green in isolation
• Smoke-tested against the published artifact (publishToMavenLocal) in a standalone lab project running a real Flamingock pipeline end to end

[dieppa]

Ran the tests and mutation-tested the filter. LGTM — a faithful mirror of the sync implementation, with the reactive plumbing done the way the rest of the module does it.

What's right:

• Survivor identified by _id, not by attribute matching — the lesson from #967 carried over rather than re-derived.
• entries.size() == 1 early return is correct here. Unlike Couchbase, MongoDB doesn't rekey, so a single document genuinely needs no work. Good that the Couchbase shape wasn't copied.
• Nothing is written, so clause 5 holds trivially and the survivor keeps its _id — with a test asserting it.
• PublisherSync on every driver call, matching MongoDBReactiveAuditRepository.
• The javadoc points at the sync version instead of duplicating the reasoning. Right call.

24 tests pass (9 unit + 15 conformance). I inverted the _id exclusion and dropped the changeId scope — both caught.

Two non-blocking comments inline.

One thing worth knowing but not this PR's problem: PublisherSync.await has no timeout, so a publisher that never completes blocks forever. That's pre-existing and applies to every reactive read/write in the module — but compaction runs inside the stage lock, so a hang would hold the lock.

For the record, a gap inherited from the sync PR: no test covers two documents with the same winning state across different executions (exec-1/APPLIED + exec-2/APPLIED). With _id-based identification that's structurally safe, so not a bug — and the shared seed can't express it today. If it ever gains that scenario, both Mongo stores benefit.