ZJIT: Do less GC work for ProfiledType

[tekknolagi]

Stats seem to show we're spending a lot more GC time than YJIT and a lot of that is due to profile information. So I want to investigate how many profiles contain objects that actually "need" to be GC-handled vs how many don't (the Integer class probably won't go away, etc). And then we can re-resign how some of this works to a) retain less and b) mark less

[tekknolagi]

For e.g. lobsters

plum% WARMUP_ITRS=0 MIN_BENCH_ITRS=20 MIN_BENCH_TIME=0 ruby --zjit --zjit-stats benchmarks/lobsters/benchmark.rb 2>&1 | grep mark_
mark_builtin_class_count:                                        495,883
mark_other_class_count:                                          743,184
plum%

Additionally,

Globally, 1972 distinct, 1238963 total classes marked in JIT code

per GC.

So we should just have a single array that gets marked once per JITstate mark and not mark at all per ISeq. Good idea @alanwu-shopify-inc