kvserver: clean up after EndTxn even if client disconnects

[erikgrinaker]

The main transaction cleanup (record and intents) is done after the Raft entry has been applied:

cockroach/pkg/kv/kvserver/replica_write.go

Lines 222 to 228 in c81a2a3

	if len(propResult.EndTxns) > 0 {
	if err := r.store.intentResolver.CleanupTxnIntentsAsync(
	ctx, r.RangeID, propResult.EndTxns, true, /* allowSync */
	); err != nil {
	log.Warningf(ctx, "%v", err)
	}
	}

However, if the client context is cancelled during Raft processing, we skip this cleanup:

cockroach/pkg/kv/kvserver/replica_write.go

Lines 291 to 297 in c81a2a3

	case <-ctxDone:
	// If our context was canceled, return an AmbiguousResultError,
	// which indicates to the caller that the command may have executed.
	abandon()
	log.VEventf(ctx, 2, "context cancellation after %0.1fs of attempting command %s",
	timeutil.Since(startTime).Seconds(), ba)
	return nil, nil, roachpb.NewError(roachpb.NewAmbiguousResultError(ctx.Err().Error()))

We should try to do this cleanup even when the context is cancelled.

As noted in #60585 (comment), this problem is exacerbated by rollbacks following client disconnects during in-flight txns or DML statements getting a 3-second context timeout -- if the request doesn't make it through RPC and Raft in that time, it won't be cleaned up.