kvserver: rangefeed txn pusher barrier may cause resolved timestamp stalls

[erikgrinaker]

In #117612, we added a barrier command to flush the rangefeed's Raft pipeline when pushing aborted transactions, as a fix for #104309.

The barrier command is marked isUnsplittable to prevent it from spanning range boundaries. Unfortunately, the DistSender enforces this constraint based on its range cache, which can be stale. Moreover, it will never attempt to refresh its range cache in response to this.

If a rangefeed runs on a follower replica after a range merge, the local DistSender's range cache may be stale, and contain the pre-merge range descriptors. When the barrier command is submitted spanning the entire merged range, it will be continually rejected by the DistSender, until some other request happens to trigger a cache refresh. This can prevent the rangefeed's resolved timestamp (and thus checkpoints) from advancing, similarly preventing the changefeed's frontier (or watermark) from advancing, logging the following error:

pushing old intents failed: range barrier failed, range split

Rangefeed events will still be emitted as usual, and garbage collection will be prevented by CDC protected timestamps, allowing the rangefeed to recover if it is restarted.

Seen in #119333.