perf(scan): fuse validate_brackets into NEON scanner

[membphis]

Summary

Eliminate the separate validate_brackets pass in the NEON scanner by carrying a depth stack inline during emit. This mirrors the scalar scanner's fused scan_and_validate approach.

Changes

• Add emit_bits_validate() that validates brackets while emitting structural offsets
• Add scan_tail_validate() for the scalar tail with inline validation
• Gate emit_bits, validate_brackets, scan_emit_resume with #[cfg] for AVX2-only (they remain used by the AVX2 scanner)
• Add dense-100k bench scenario (46% structural density) to measure fusion impact

Performance

Measured improvement: ~3% (within noise)

Profile analysis predicted validate_brackets consuming ~30% of scan time on structure-dense payloads. However, benchmarks show only ~3% end-to-end improvement even on a 46% structural density payload.

This confirms issue #25's analysis: the per-emit buf[pos] lookup to determine bracket type offsets the savings from eliminating the separate pass.

Payload	Structural Density	Old (ops/s)	New (ops/s)	Change
dense-100k	46.57%	9,360	9,677	+3.4%
small (2KB)	12.72%	831,947	848,464	+2.0%
medium	0.12%	362,319	366,300	+1.1%
100k (multimodal)	0.08%	278,552	280,899	+0.8%

Why the limited gain:

1. Each emitted structural character now requires a buf[pos] random memory access
2. This cost roughly equals the eliminated validate_brackets pass
3. String-heavy payloads (multimodal, base64) have <0.1% structural density, so validate_brackets was never a bottleneck there

Value of this change:

• Cleaner single-pass architecture (matches scalar scanner design)
• Marginal improvement on structure-dense workloads
• No regression on any workload

Testing

• cargo test --release — NEON gate
• cargo test --release --no-default-features — scalar unchanged
• cargo test --features test-panic --release — FFI panic barrier intact
• make lint — clippy clean
• make bench — no regression, ~3% improvement on dense payloads

Closes #25

[membphis]

Closing: benchmark shows only ~3% improvement (within noise), confirming that per-emit buf[pos] lookups offset the eliminated pass. Not worth the added complexity.