Fold identical functions in the bindings on every platform, not just Linux

[Fedr]

--icf=all has been Linux-only since it was added, with the note "until lld-link (Windows) and ld64.lld (macOS) get their own measurements". Those measurements are now in, from #6731: every macOS leg linked the module with ICF and ran the full Python sanity suite on the result.

leg	baseline	with --icf=all	sanity
macOS x64 Release	45,815,344	39,606,312 (-13.6%)	pass
macOS arm64 Release	30,881,104	24,925,216 (-19.3%)	pass
macOS arm64 Debug	32,046,672	25,803,472 (-19.5%)	pass

__text alone drops 12.8% on x64 and 20.1% on arm64. arm64 folds more, as expected: fixed-width instructions make more of the ~190k near-duplicate template instantiations byte-identical.

What changed

Only the IS_LINUX guard around the flag, replaced by one that excludes the two platforms that cannot use it. It stays inside the MODE=release preset, because that is where -flto=thin comes from and lld's LTO codegen is what emits the per-function sections ICF folds; a MODE=none build has nothing to fold on any platform.

• Emscripten — wasm-ld has no ICF. This was covered implicitly before (IS_LINUX is cleared when targeting Emscripten); now it is explicit.

• Windows — nothing to add: lld-link already folds by default. Measured on the wheel build, which unlike build-test-windows.yml uses this preset:

  wheel module	baseline	/opt:icf	/opt:noicf
  mrmeshpy.pyd	56,210,432	56,210,432	63,037,440 (+12.1%)
  mrcudapy.pyd	816,640	816,640	876,032 (+7.3%)

  /opt:icf is a no-op because ICF comes with /opt:ref in non-debug links, and disabling it costs 6.8 MB. The GNU spelling never reached the linker's logic there either — lld-link: warning: ignoring unknown argument '--icf=all', byte-identical output — so passing it on Windows would only add a warning.

Not included

-fomit-frame-pointer was measured in the same run and is not usable on macOS: it grows the module by 12-15% (Apple's ABI expects a frame pointer, so dropping it pushes functions out of the compact-unwind encodings into much bulkier DWARF entries) and it breaks the sanity suite outright — libc++abi: terminating due to uncaught exception of type pybind11::stop_iteration, i.e. unwinding stops finding handlers. Alignment padding is already nil under -Oz (477 bytes in a 34.7 MB __text), so there is nothing there either.