Optimize current RF chroma shift pass

[JunliangRen]

Summary

• fuse the first float32 quantization into the current RF chroma roll/DC pass and vectorize the final centering conversion
• add an independent frozen oracle with AVX-required and AVX-disabled CI gates, including heap-wrap and IEEE special-value coverage
• refresh the English, Chinese, and Japanese performance tables and detailed evidence from one 60-run matrix

Compatibility and performance

• all 60 Exact/IPP-fast matrix runs were deterministic within and across default/1/5/10/20 worker settings
• luma, chroma, raw JSON, stdout, normalized stderr/logs, ordered fileLoc, and field count remained identical
• the 32,768-sample kernel improved from 28.025 ms to 16.166 ms over 1,000 calls (42.317%)
• three 1,000-frame Exact current --threads 20 A/B pairs retained all compatibility surfaces; paired median wall/CPU gains were 0.5385%/0.6860%

Validation

• Release build: 0 warnings, 0 errors
• xUnit v3: 1,615 discovered, 1,612 passed, 3 expected environment skips
• AVX-required focused test: 1/1 passed
• AVX-disabled chroma reduction tests: 4/4 passed
• localized README tests: 2/2 passed
• local independent review findings addressed