crosscluster/logical: transactional LDR emits one duplicate endTime checkpoint per remaining KV

**Describe the problem**

The type comment says the synthetic `endTime` checkpoint is deduplicated by
`maybeEmitCheckpoint`, but `processPastEndTime` calls `emitCheckpoint`
unconditionally — one checkpoint per remaining KV after cutoff, each broadcast to
every applier, precisely while the convergence loop is trying to wind down.

**Impact**

Checkpoint storm during convergence teardown (extra broadcasts per remaining KV);
inefficiency during the sensitive wind-down phase.

**Code references**
- [txnfeed/merge_feed.go:222](https://github.com/cockroachdb/cockroach/blob/master/pkg/crosscluster/logical/txnfeed/merge_feed.go#L222)

**Suggested fix**

Call `maybeEmitCheckpoint` (the first emission still fires).

Epic CRDB-65552

Jira issue: CRDB-65608

Start in pkg/crosscluster/logical/txnfeed/merge_feed.go at processPastEndTime and inspect maybeEmitCheckpoint alongside the referenced line. Verify the existing checkpoint behavior and relevant txnfeed tests, then confirm that the first synthetic endTime checkpoint is emitted while subsequent remaining KVs do not cause duplicate broadcasts.

Written by the indexing model from the issue text.