Vectorize JSON escaping

[xosnrdev]

Summary

Vectorize JSON byte escaping with NEON, AVX2, and AVX-512BW; eliminate per-entry heap allocation in SSTable reads; replace linear memtable scan with binary search.

Latency result (x86_64 AVX2)

Scalar is the byte-at-a-time lookup table from master. Vectorized is the SIMD scanner on this branch.

Workload	Size	Scalar	Vectorized	Speedup
clean (0% escapes)	64 B	46.2 ns	20.6 ns	2.2x
	256 B	146.8 ns	23.1 ns	6.4x
	1 KiB	585.8 ns	83.2 ns	7.0x
	4 KiB	2.18 µs	160.9 ns	13.5x
	16 KiB	8.62 µs	455.9 ns	18.9x
mixed (2% escapes)	64 B	48.3 ns	26.5 ns	1.8x
	256 B	153.4 ns	63.3 ns	2.4x
	1 KiB	607.5 ns	270.4 ns	2.2x
	4 KiB	2.38 µs	933.7 ns	2.5x
	16 KiB	9.10 µs	3.61 µs	2.5x

Core Components

• simd crate (new): write_escaped_json_bytes dispatches to find_next_escape_avx512bw, find_next_escape_avx2, find_next_escape_neon, or find_next_escape_fallback based on arch/CPU features
• FIND_NEXT_ESCAPE_DISPATCH: self-patching static mut function pointer -- CPU detection runs once, subsequent calls pay one indirect jump
• SstableReader: gains data_buf: Vec<u8> field; reused across entries via resize instead of allocating per read
• ReadSnapshot::get: linear scan over memtable_entries replaced with binary_search_by_key

Platform Support

• aarch64: NEON 16-byte lanes via vcltq_u8 / vcgtq_u8 / vceqq_u8, bitmask extracted with vpadd_u8 + vreinterpret_u16_u8
• x86 / x86_64: runtime dispatch via is_x86_feature_detected!; AVX-512BW processes 64 bytes/iter using __mmask64 OR, AVX2 processes 32 bytes/iter using vpmovmskb; LLVM lowers both to a vpaddb + vpminub/vpcmpltub saturation range check
• all: SWAR fallback processes 8/4/2/1 bytes per step using the Falk Hüffner has-zero-byte identity

Helpers

• find_next_escape_raw: selects arch path at compile time--falls through to find_next_escape_fallback on unsupported targets

Tests

• Add invariant test for simd::json