JIT: Move executor to a register

[diegorusso]

With this change we put the executor in one of the register so that _MAKE_WARM and _CHECK_VALIDITY don't need to patch the executor and they become 1-2 machine instructions (depending on the platform)

On AArch64 there is a small uplift of 1% and in x86_64 the net effect is about 0% (even though we are using a caller saved register)

[Fidget-Spinner]

Hmmm, I would think that maybe instead of this, we allocate one more register to TOS caching instead for AArch64? I think that would probably provide bigger gains?

[diegorusso]

Hmmm, I would think that maybe instead of this, we allocate one more register to TOS caching instead for AArch64? I think that would probably provide bigger gains?

The executor should be hotter than the fourth TOS item and it doesn't ever need to be spilled. This change also allow further optimisations in case we need to put things accessible via the executor.

[Fidget-Spinner]

Ok let me just run the benchmarks on the Meta runners for x86-64. Will approve after this.

[Fidget-Spinner]

There's zero change on x86-64 https://github.com/facebookexperimental/free-threading-benchmarking/blob/main/results/bm-20251222-3.15.0a3%2B-92c1f91-JIT/bm-20251222-vultr-x86_64-diegorusso-executor_register-3.15.0a3%2B-92c1f91-vs-base.md

Though nbody slows down a little, probably due to a little more register pressure.