perf: SIMD-accelerate JSON string paths (from_json/to_json)

[khanaffan]

Summary

Adds 128-bit SIMD (ARM NEON + x86 SSE2, with a scalar fallback) to the two byte-at-a-time string loops in the msgpack⇄JSON conversion — the one place in this extension where raw byte throughput actually dominates.

• mpJpParseString (msgpack_from_json): SIMD-scan for "/\, then bulk-copy the safe run in a single append.
• mpJsonEscapeStr (msgpack_to_json / msgpack_pretty): SIMD-scan for "/\/control bytes, then jump straight to the next byte that needs escaping.

New helpers mpScanStr/mpScanEsc:

• Use full 16-byte loads only (i+16<=len) so we never read past the buffer; the short tail is handled scalar.
• Use only the always-available baseline ISA (NEON on arm64, SSE2 on x86-64) → no runtime CPU dispatch.
• A cheap first-byte guard avoids any regression on escape-dense input.
• -DMSGPACK_DISABLE_SIMD forces the portable scalar path.

Why not extract/set?

MessagePack is a length-prefixed TLV format, so element navigation (mpSkipOne/mpLookup) is a serial pointer-chase that can't be vectorized; key/buffer ops already use libc SIMD memcmp/memcpy. Those paths are intentionally left untouched.

Measured gain

arm64/NEON, Release, ~4 KB strings, A/B vs MSGPACK_DISABLE_SIMD:

case	scalar	SIMD	speedup
from_json, no escapes	~2670 ns	~579 ns	~4.6x
to_json, no escapes	~2590 ns	~510 ns	~5.0x
to_json, all-escapes (worst)	~8500 ns	~8500 ns	~1.0x (parity, no regression)

Small strings are unchanged (bound by SQLite call overhead) — as expected.

Correctness

• 16/16 ctest pass; fuzz corpus pass.
• 15 hand-crafted boundary cases (escapes at 16-byte chunk offsets 15/16/17, control bytes, NUL, tails, empty): SIMD output byte-identical to scalar + round-trip OK.
• 6000-case deterministic differential fuzz (random lengths/offsets/specials): SIMD output byte-identical to scalar.
• x86-64 SSE2 branch cross-compiles clean; MSVC-safe (_BitScanForward).

Changes

• src/msgpack.c — SIMD scanners + wired into the two JSON string loops.
• tests/bench_msgpack_vs_json.c — large-string benchmark rows (plain + all-escapes).