feat: image-adaptive codec knob presets from coefficient benchmarks

## Summary
coefficient's `knob_presets.rs` has empirically-derived per-codec config recommendations based on image classification (from 1125 A/B measurements on CID22). Move these to zencodecs where codec dispatch lives.

## Key findings
- JPEG: 4:4:4 chroma wins +0.76 zensim on HighDetail, hurts Text/Uniform
- AVIF: 10-bit worth it only for Uniform (gradients), e8 sweet spot
- WebP: sharp_yuv BANNED (−17 to −26 zensim regression in zenwebp)
- JXL: gaborish always on (−6-19% BPP without it), patches only for text

## Dependencies
Requires zencodec `analysis` feature (imazen/zencodec#13) for `ImageType` classification.

Start by reading coefficient's knob_presets.rs and the zencodecs codec-dispatch code, then confirm zencodec#13's analysis feature and ImageType API. Port the empirically derived per-codec presets and validate them against the listed CID22 image-classification findings, including the WebP sharp_yuv prohibition. Done when dispatch applies the appropriate presets for JPEG, AVIF, WebP, and JXL.

Written by the indexing model from the issue text.