[ move WRITE_ONCE(gc_in_progress, true) into the __unix_gc() work function and drop it from unix_gc(). ]

Dropping the set from unix_gc() and doing it in the work function is fine
upstream, but only because later refactors made wait_for_unix_gc() gate
flush_work() on unix_graph_cyclic_sccs. On these trees it still gates on
gc_in_progress, so this brings back a window where the over-limit throttle
stops waiting for GC.

Can you respin keeping gc_in_progress set in unix_gc() before queue_work()?
The code is identical across all three, so one version covers them.

--
Thanks,
Sasha