fix rare GC segfault coming from bug in mark-loop (alternative to 53351)

[d-netto]

Should fix #52757, #53026 and #53350.

Removes some of the awkwardness from #53351 but does require waiting for all threads to join at the end of mark-loop.

[vtjnash]

I think #53355 should fix all 3 of those now (there appeared to be actually 3 different issues on review of the code), without needing the full pthread_barrier-like construct added here, which might slow down the master thread slightly more (since it only needs to wait for its own turn to acquire the lock, rather than waiting for every thread to get a turn to acquire the lock)

[d-netto]

SGTM. Will let it be superseded by #53355. We can re-open if needed.