Multiple Workers to Concurretly Execute Large ROW-format Transactions

For workloads with mixed transaction sizes (small/large), a bottleneck
is often group-commit, where future small transactions must wait for a
previous large transaction to commit, before they can commit.

With binlog_format=row, we know the data ahead of time. This ticket
covers the idea of splitting a large ROW-format transaction into
separate pieces, where each can be applied concurrently by multiple
threads. Specifics need to be researched.

This would would likely be blocked by MDEV-37587, as I imagine this
project would require some disconnect between slave threads and
transactions. I.e., perhaps some task-based execution model, where
applying row event data is considered a type of task that any idle
thread could pick up.