perf(atomsplit): fuse o200k letter case-split + 32-wide ASCII classify

[ArthurZucker]

Two byte-exact perf changes to atomsplit, both isolated to the scalar o200k FSM and the NEON classify kernel. No behavioural change — o200k_parity and the classify SIMD-vs-scalar parity tests pass, 0 warnings.

1. o200k letter case-split: one pass instead of two

fsm_o200k walked every [\p{L}\p{M}]+ run twice — letter_end() to find the run end, then emit_o200k_letters/o200k_letter_match re-scanned the same bytes for the case split. Since the tags already carry exact case (UpperLetter/LowerLetter/caseless), the run end can be discovered inline during the case split:

• fold letter_end into the phase-A [UC]* / phase-B [LC]+ scans (each stops at a non-member = run end);
• unroll those scans 16 tags/chunk with get_unchecked, same shape as run_end (a 16-wide scan already skips cont spans wholesale — a 48-wide variant was measured and lost);
• resolve the give-back last-caseless target by a backward hop only on the no-lowercase exit, so the forward scans stay pure skips (no per-char tracking).

FSM throughput (ns/byte, min-of-7, byte-exact ✓ vs onig):

lang	before	after	Δ
English	2.163	1.602	−26%
French	1.936	1.437	−26%
Russian	1.243	0.761	−39%
Greek	1.316	0.824	−37%
Arabic	1.539	1.104	−28%
Thai	1.076	0.636	−41%
Chinese	1.154	0.759	−34%
Japanese	1.126	0.764	−32%
Korean	1.746	1.371	−21%

Every corpus now beats logos on the full (classify + fsm) pipeline; the previous Chinese/Thai ties are gone.

2. classify: 32-wide ASCII fast path

The NEON kernel loads 32 bytes to produce 16 tags (it needs the b1/b2 lookahead), so non-ASCII bytes were effectively loaded twice. The ASCII fast path now peeks 32 bytes at once — one bounds check + one vmaxvq reduction covers two 16-byte stores — done in place. The multibyte body is completely untouched (same code, same comments). Real text is ~half ASCII (spaces / digits / punctuation), so this fires constantly.

classify throughput (ns/byte, byte-exact ✓ vs classify_scalar):

lang	before	after	Δ
English	0.153	0.085	+45%
French	0.252	0.218	+14%
Russian	0.384	0.344	+10%
Chinese / Korean / Thai	—	—	~neutral

No regression on non-ASCII scripts. (A full 32-wide multibyte driver was prototyped too — it gains Chinese +17% but regresses 2-byte scripts −8 to −14%, so only the clean ASCII-widening is kept.)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.