kvclient: corrrectly handle WriteTooOld error in read committed in txnWriteBuffer

[stevendanna]

The buffered writes project decomposes some requests into a locking get and a buffered write request. This scheme requires the locking get to provide isolation through commit time. At the moment we use unreplicated locks for this case which can still be lost in some cases.

If a write sneaks in underneath us we will still detect this in a read-committed transaction when we go to flush the batch. But, we need to carefully handle the resulting WriteTooOld error in this case. Unlike a normal RC transaction, a write buffered RC transaction cannot arbitrarily move the WriteTimestamp of the request forward because this could result in what would appear to users as single-statement write skew, which we don't currently allow.

Jira issue: CRDB-48875