poor XA performance on non-binlogging parallel slave

For performance improvement, which is about the same as in unfixed 10.6 base of MDEV-31949 or MDEV-33668, the non-binlogging parallel slave requires a different approach for effective grouping of parallellly executed XA:s.
Namely to make use of / specialize MDEV-21534 framework for slave ordered xa prepare/commits.

Specifically parallelly executing XA-PREPARE:s should not rush to prepare themselves durably in engines logs, but rather be cognizant of a leader transaction that must eventually wrap them all together in one group to durably prepare.