Build smaller turbo binaries - #6552

Every pointer in a constant table is a load-time relocation: 24 bytes of
.rela.dyn on ELF plus its 8-byte slot. The generated reg::Arg and
reg::Sig records were 27,157 of the Linux .so's 30,374 relocations.

Each generated header now has one string table and one parameter table
(namespace sigtab); the records hold 16-bit offsets and indexes into them,
and reg::Sig carries the two tables as template arguments, which compile
to PC-relative addresses at the registration sites, so cls.method(sigs::x,
...) stays as it was. Linux arm64 (GCC 11.4, -O2): relocations 30,374 ->
2,995, .rela.dyn 736 KB -> 78 KB, .data.rel.ro 506 KB -> 27 KB.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qi7GRGMn9ZuY2jSc97wqBF

They run once per process at module startup; ZEND_COLD makes the compiler
optimise their bodies for size and move them out of the hot text. The
method handlers they register are separate functions whose addresses
escape, and keep their optimisation.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qi7GRGMn9ZuY2jSc97wqBF

When the OPcache settings are not the ones TurboProcessRestarter wants
(OPcache off for CLI in the CI images, validated timestamps on a default
Homebrew PHP), PHPStan re-executes itself and a command-line
-d extension= does not survive: the run measures PHP alone while
php -d extension=... -r still reports the extension's version. On
src/Type src/Analyser --debug: 40.0 s user with -d, 17.2 s through
PHP_INI_SCAN_DIR.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qi7GRGMn9ZuY2jSc97wqBF

The shipped .so had grown to ~12 MB per platform and PHP version, 369 MB
across the 38 binaries in phpstan/phpstan. Measured on Linux arm64 (GCC
11.4, make pgo), stripped:

- strip the symbol tables in CI (make strip, before signing and before the
  tests, so they run the shipped binary): 12.0 MB raw -> 8.9 MB
- no unwind tables: the extension never throws and never walks its own
  stack, yet -fno-exceptions alone still leaves .eh_frame (0.8 MB)
- per-function sections with --gc-sections / -dead_strip
- LTO under GCC: 3.9% faster on top of PGO and 15% smaller; clang's full
  LTO was 1.2% faster but 9% larger, so macOS builds stay without it
- main.cpp supplies the terminate handler in the static-libstdc++ build,
  keeping libstdc++'s demangler out of the link

With the pointer-free signatures and cold registration, the Linux .so
goes from 12.0 MB to 8.3 MB raw and 6.3 MB stripped, 4.1% faster than
today's CI build; macOS from 10.5 MB to 7.1 MB stripped at unchanged
speed. -Os was measured and rejected: a third smaller than -O2, but 10.4%
slower. PGO keeps its place: 4.0% faster on Linux, 3.6% on macOS. The
sanitize target restores the unwind tables its stack traces need.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qi7GRGMn9ZuY2jSc97wqBF