Put each function in its own section so ICF can fold more

[Fedr]

Follow-up to #6547. --icf=safe can only fold what the compiler put in its own section: template instantiations and inlines get COMDAT sections automatically, but plain out-of-line functions share one .text per translation unit and are invisible to folding. -ffunction-sections gives each of them its own section.

Same gate as the ICF flag — inside the MESHLIB_HAVE_LLD block, non-Debug only — so nothing changes for GCC rows, macOS, emscripten or Debug builds. Restricted to C/C++ via $<COMPILE_LANGUAGE:C,CXX>, because add_compile_options would otherwise hand -ffunction-sections to nvcc, which does not accept it. -fdata-sections is deliberately absent: lld's ELF ICF folds executable sections only.

Sizes: −2.0% once stripped, on both arches

Release / Clang 21 / lld, all 14 .so, baseline = master (--icf=safe already in):

	baseline	-ffunction-sections	delta
x64, unstripped	62,285,048	61,363,448	−921,600 (−1.48%)
x64, stripped	46,020,984	45,099,384	−921,600 (−2.00%)
arm64, unstripped	64,009,152	65,737,856	+1,728,704 (+2.70%)
arm64, stripped	46,360,888	45,443,384	−917,504 (−1.98%)

The arm64 growth in unstripped builds is entirely symbol tables — per-function sections mean per-function section symbols:

arm64 libMeshLibC2.so	baseline	variant	delta
.symtab	2,476,176	4,485,936	+2,009,760
.strtab	4,250,085	4,501,305	+251,220
.gcc_except_table	1,087,764	195,492	−892,272

.symtab/.strtab are not SEC_ALLOC: never loaded at runtime, and removed by strip. Everything we ship is stripped — wheels via #6525, the AppImage by linuxdeploy (which strips unless NO_STRIP is set) — so the +2.70% never reaches a user.

Where the win comes from

Not .text. Summed over all .so (x64 / arm64):

section	delta
.gcc_except_table	−845,124 / −916,720
.text	−55,192 / −58,592
.eh_frame + .eh_frame_hdr	−18,292 / −20,348

The folding is mostly exception-handling tables. libMeshLibC2.so alone drops .gcc_except_table from 1,045,892 to 220,576 — −79% — while its .text is byte-identical: the generated C bindings are thousands of near-identical try/catch wrappers, so their LSDAs are duplicates even where the surrounding code is not.

Caveat before merging

scripts/distribution_vcpkg.sh does not strip (cmake --install without --strip), so the linux-vcpkg tar.xz is the one artifact that would carry the extra symbol tables. Its compressed delta is unmeasured and could go either way: the added .symtab and the removed .gcc_except_table both compress well. Two ways to settle it — measure the package both ways, or strip in the packaging script, which makes the question moot and shrinks that artifact regardless of this PR.

Testing

Linux legs only; Windows and macOS disabled as unreachable behind UNIX AND NOT APPLE. All four linux-vcpkg legs green across every measurement run.