Race condition in trx_purge_rseg_get_next_history_log()

In MDEV-15912, I made a mistake. The function trx_purge_rseg_get_next_history_log() would release the undo page latch by committing the mini-transaction. After that, we would wrongly read the page frame to assign a value of purge_sys.rseg->needs_purge.

This was caught by ASAN heap-use-after-poison in a stress test, presumably related to buffer pool resizing.