ext/standard: Optimize var_export - #23189

I measured this on master c65321c (dc5f795 merges with a conflict in UPGRADING only) with GCC 13.3, GCC 14.2, clang 18.1 and clang 20.1 on an i7-13700H, built with the flags of the official Docker image. On 32-bit builds the escaped length overflows and var_export() crashes; details and a replacement are in my review comment on var.c.

Change in CPU cycles per var_export($value, true) call with the PR, relative to master (negative = fewer cycles, faster); median of five alternating runs:

Long strings get slower. Strings, arrays, config arrays and objects of small and medium size take 15–68% fewer cycles, since the temporary strings and spprintf() are gone. Long strings take more: the PR reads every byte twice, once to count and once to write, and both loops test for ', \ and NUL separately, so a 64 KB string with ' or \ after one word in eight runs 23–28 instructions per byte where master runs 8–9 (php_addcslashes() looks each byte up in a table once, php_str_to_str() finds NULs with memchr()).

The one-pass version in my review comment (a byte class table, plain runs appended with smart_str_appendl()) is faster than master in every case I measured except a single int (−2 … +2%, within the −1 … +2% of master against master), on all four compilers:

Correctness: output identical to master with the PR and with that version, on all four compilers, under ASan/UBSan and on a 32-bit build, for every byte value as value and key, all strings up to 4 bytes over a, ', \ and NUL as values and keys, and as property names where valid (not empty, not starting with NUL), long runs and mixes of them, nested arrays and objects with visibility and enums, and 200,000 random nested values; the harness catches six deliberate mutations of the PR (11 → 10 per NUL, a skipped escape, one space less, NUL escaped in property names, and others).

Minor: buffer_append_spaces() takes an int. A level of −1 through php_var_export_ex() (it is PHPAPI) gives array elements a width of 0, where %*c still wrote one space, and a level of −2 or below gives a negative width that becomes SIZE_MAX in memset(). In-tree callers pass 1 or more.

On buffer_append_spaces() (the question in the description): as a plain static function it is inlined by clang 18 and 20 and kept out of line by GCC 13 and 14 (7 call sites). Arrays and objects gain about the same with both (1,000 ints: −46 … −49% with GCC, −45 … −47% with clang), so leaving it without inline is fine.

Full tables, scripts and data: https://github.com/ArtUkrainskiy/php-src-bench/tree/main/reports/pr-23189-var-export