[BENCH, do not merge] Measure -fomit-frame-pointer and --icf=all on the macOS bindings

[Fedr]

Benchmark only, not for merge. Measures the two size levers that are left for the bindings after -Oz and ThinLTO, on the macOS legs.

Why these two, and not the usual suspects: I scanned the published mrmeshpy.so of 3.1.3.429 and -Oz already leaves 477 bytes of alignment padding in a 34.7 MB __text, so -falign-functions and friends have nothing to recover. What the scan did find is 202,218 push rbp; mov rbp,rsp prologues in the macOS x86_64 build against 3 in the Linux one — Apple's Clang keeps frame pointers on x86-64, Linux Clang omits them at -O2+ — worth roughly 1.2 MB with the epilogues, plus whatever freeing rbp saves in spills.

--icf=all is restricted to Linux with the note "until lld-link (Windows) and ld64.lld (macOS) get their own measurements", where it removed 11% of the module. This is that measurement for macOS.

The step prints file size and __text for four variants:

variant	how
baseline	the module the job just built
--icf=all	re-link of the same objects, link-only flag
-fomit-frame-pointer	fragments recompiled, then linked
both	re-link of those fragments with ICF

Only one recompile round is needed, so this adds roughly 15-25 min per macOS leg. Every non-macOS platform is disabled by label. I will close this PR and delete the branch once the numbers are in.

Note that neither flag touches the actual x86-64 vs arm64 gap (__text 34.7 MB vs 21.4 MB on macOS, 34.4 vs 21.9 on Linux) — that is instruction encoding and register count, and no flag fixes it.