Algorithmic improvements for LateLowerGC pass (#47634)

On a particularly pathological case I'm looking at (~100k calls,
each a safepoint, similar number of allocations), I'm seeing
LateLowerGC pass take the majority of the middle end time
(around 80s out of 100s). This PR goes through and improves
some of the algorithmics and data structure choices and gives
a roughly 10x improvement on this particular pathological case
(down to about 7.6s on my machine). That still represents about
1/3 of the total middle end time, which I'm not happy about, but
perhaps that is reasonable given how pathological my particular
test case is. For comparison, register allocation (which needs
to solve a somewhat similar problem) on this IR takes about 20s.

Author: Keno