feat(models): port Hunyuan-VL (ViT + perceive merger + XD-RoPE decoder)

[inureyes]

Summary

Adds Hunyuan-VL (hunyuan_vl, e.g. HunyuanOCR), Tencent's vision-language family.

• Vision tower (src/vision/encoders/hunyuan_vl.rs): a ViT over raster-ordered 768-wide patch rows. The per-patch conv embedding is folded into a linear at load; a learned position-embedding table (CLS row skipped) is bilinearly interpolated to each image's grid with the upstream truncation-toward-zero corner math; blocks run full attention over the packed sequence (matching upstream, which attends across images) with exact-GELU MLPs. The perceive PatchMerger per image applies RMSNorm, a stride-2 conv pair over the raster grid, a learned image_newline column, a linear to the decoder width, and learned image_begin / image_end rows: mh * (mw + 1) + 2 output rows per image.
• Placeholder contract: the merger's row count IS the prompt placeholder count, so the runtime reuses the count-based DeepSeek-OCR placeholder expansion (insert_deepseekocr_image_tokens) with no framing tokens; the framing is carried by the features.
• Decoder (src/models/hunyuan_vl.rs): the Hunyuan dense stack (per-head Q/K RMSNorm applied after the rotation, DynamicNTK-alpha rope base theta * alpha^(d/(d-2))) with XD-RoPE at prefill: 4D [P, T, H, W] position ids split across the head_dim / 2 frequency dims per xdrope_section ([16, 16, 16, 16]). P stays sequential everywhere; inside an image run W cycles 0..=mw per row (the newline slot takes w = mw), H repeats per row, T is the image index. A unit test proves the text-only degeneration to the standard concat-half rotation (fast_rope, traditional = false); decode uses sequential positions (rope_deltas = 0) through the shared MRopeState plumbing.
• Processor (src/vision/processors/hunyuan_vl.rs): smart resize to multiples of 32 in [min_pixels, max_pixels] (defaults 512^2 / 2048^2, preprocessor_config.json overrides honored), bilinear, CLIP normalization, plain raster patch rows.

Validation

Real checkpoint hadeseus/HunyuanOCR-mlx-4bit (~1B; text 4-bit, vision tower bf16):

• Text-only: "The capital of France is Paris." (120 tok/s)
• Character-exact OCR of a rendered image: "MLXCEL RUNTIME Invoice No. 2026-0705"
• Shapes image: blue square upper right, green triangle lower, colors and positions correct.

8 co-located unit tests (XD-RoPE degeneration, alpha base scaling, processor grid/rows/count) and 4 real-model parity tests pass. Release build, cargo fmt, and cargo clippy --all-targets -- -D warnings are clean.

Out of scope

Video inputs, multi-image XD-RoPE beyond the per-image grid pattern above, and server per-sequence MRoPE binding beyond the PaddleOCR-VL fallback-slot scope.

Closes #536