YJIT: port call threshold logic from Rust to C for performance

[maximecb]

This way the C compiler can potentially inline the YJIT call threshold logic, and we avoid allocating/touching the ISEQ payload object or needing to take the VM lock to do call threshold accounting.

[k0kubun]

Benchmarked this with --yjit-cold-threshold=0 and fib (since the result of 30k_methods was unreliable for some reason 🤔). It's looking good.

before: ruby 3.3.0dev (2023-10-11T14:16:39Z master cc311e1c45) +YJIT [x86_64-linux]
last_commit=[ruby/irb] Avoid locking the debug UI to a single thread (https://github.com/ruby/irb/pull/725)
after: ruby 3.3.0dev (2023-10-11T21:16:38Z yjit_c_threshold_l.. ba3587b89d) +YJIT [x86_64-linux]
last_commit=Exclude yjit_calls_at_interv from RJIT bindgen

-----  -----------  ----------  ----------  ----------  -------------  ------------
bench  before (ms)  stddev (%)  after (ms)  stddev (%)  after 1st itr  before/after
fib    237.9        0.1         226.7       0.2         1.06           1.05
-----  -----------  ----------  ----------  ----------  -------------  ------------

[maximecb]

Oh wow, that much speed difference! I'm surprised, but seems it was really worth it!

We should expose rb_yjit_compile_new_iseqs() as a Rust-side global next 🤔