28911: release-2.1: kv: lie better about commits that are really rollbacks r=andreimatei a=andreimatei

cc @cockroachdb/release 

Backport #28872

When a client tries to commit a txn that has performed writes at an old
epoch but has only done reads at the current epoch, one of the
TxnCoordSender interceptors turns the commit into a rollback (for
reasons described in the code).
This patch completes that interceptor's lie by updating the txn status
upon success to COMMITTED instead of ABORTED. Since a commit is what the
client asked for, it seems sane to pretend as best we can that that's
what it got. In particular, this is important for the sql module, where
the ConnExecutor looks at the txn proto's status to discriminate between
cases where a "1pc planNode" already committed an implicit txn versus
situations where it needs to commit it itself. This was causing the
executor to think the txn was not committed and to attempt to commit
again, which resulted in an error.
I don't know if we like the ConnExecutor looking at the proto status,
but I'll leave that alone.

Fixes #28554
Fixes #28796

Release note: None

Co-authored-by: Andrei Matei <[email protected]>