feat(models): port ERNIE-4.5 MoE VL (DFNRope ViT + modality-split MoE + 3D MRoPE)

[inureyes]

Summary

Adds ERNIE-4.5 MoE VL (ernie4_5_moe_vl), Baidu's vision-language MoE, with two structures new to the tree.

• Modality-split MoE (src/models/ernie4_5_moe_vl.rs): each MoE layer holds independent text (gate / switch_mlp) and multimodal (gate_1 / switch_mlp_1) expert banks. Routing per bank is softmax in f32, top-k selected over probs + e_score_correction_bias (the bias shifts selection only), mixing weights are the uncorrected probs renormalized by max(sum, moe_norm_min). Token-type dispatch takes the text bank for text tokens and the multimodal bank at image-placeholder positions; text-only prompts and every decode step skip the multimodal bank entirely. A fused shared-experts SwiGLU is added for all tokens. Config MoE fields parse in int-or-2-list form.
• Interleaved 3D MRoPE: [T, H, W] axes assigned per frequency index (H on even, W on odd indices below mrope_section[0]+[1] = 44; T above), rotating adjacent element pairs in f32. A unit test proves the exact degeneration to fast_rope(traditional) for scalar positions. Position-id construction and per-sequence delta tracking reuse qwen_mrope_state::MRopeState (PaddleOCR-VL plumbing).
• DFNRope ViT (src/vision/encoders/ernie4_5_vl.rs): reuses the Qwen2-VL packed cu_seqlens attention, merge-window 2D RoPE, and rotation helpers; a plain 588-wide linear patch embedding (no conv reorder, no bias), quick_gelu MLP, trailing LayerNorm, no in-encoder merger.
• Variable-resolution resampler (src/vision/connectors/ernie4_5_vl.rs): 2x2 spatial fold over consecutive merge-window rows, even/odd temporal pair fold (a single image duplicates its frame), GELU MLP stacks with LayerNorms, then a Linear to the decoder width plus RMSNorm.
• Processor (src/vision/processors/ernie4_5_vl.rs): Qwen2-VL-family smart resize (factor 28, CLIP normalization), 588-wide rows in merge-window order (load-bearing for the resampler fold, unit-tested), grid_thw = (1, gh, gw).
• Runtime: reuses insert_qwen_vl_image_tokens (ERNIE's <|IMAGE_START|>/<|IMAGE_END|> frame the <|IMAGE_PLACEHOLDER|> run, end = start + 1) and the QwenVlm preparation summary; no new summary variant.

Tokenizer fix (required by this family)

decode on the SentencePiece path errored "Out of range" for ids outside the SentencePiece vocab that added_tokens_decoder declares with special: false (ERNIE's image ids: the sp vocab ends at 100295, <|IMAGE_PLACEHOLDER|> is id 100295 with special: false). The CLI swallowed that error into empty output, so image runs displayed nothing while generating correct tokens. Non-special added tokens now decode to their content strings (they are real text per HF semantics and are never skipped).

Validation

Real checkpoint mlx-community/ERNIE-4.5-VL-28B-A3B-Thinking-4bit (28B-A3B; text and resampler 4-bit, vision tower bf16):

• Text-only: thinking trace then "Paris."
• Shapes image (324 image tokens): "a red circle positioned on the top left, a blue square positioned on the top right, a green triangle below the circle and square." Colors, shapes, and positions all correct.
• Red/blue split image: "The color on the left side is red. The color on the right side is blue."
• Observed quirk: a fully uniform solid-color image is described as a checkerboard through the chat template (the same image answers "Red" through a raw prompt); structured images localize correctly, so this reads as thinking-model behavior on contentless input rather than a pipeline fault.

10 co-located unit tests (MRoPE degeneration, selection-only correction bias, MoE layer rule, int-or-2-list parsing, merge-window row order) and 4 real-model parity tests pass; the parity suite documents --test-threads=1 since two tests each load the 16 GB checkpoint. Release build, cargo fmt, and cargo clippy --all-targets -- -D warnings are clean.

Out of scope

Video inputs, the moe_use_token_type_bias training path, server per-sequence MRoPE binding beyond the PaddleOCR-VL fallback-slot scope, and grounding post-processing.

Closes #535