Possible hang in trx_undo_prev_version()

MDEV-34515 makes the InnoDB MVCC code acquire a shared purge_sys.latch twice. Recursive shared latch acquisition may cause a deadlock of InnoDB threads if another thread in between will start waiting for an exclusive latch.

Stack traces of the hang were provided earlier by elenst and again this week by mleich. With a debug patch applied, mleich was able to generate an rr replay trace that clearly shows this problem.