feat(vlm): add MiniMax-M3-VL multimodal support

[inureyes]

Summary

Add the vision-language variant of MiniMax-M3 (model_type: "minimax_m3_vl"). Depends on the MiniMax-M3 text port (#763).

Architecture specification

• Vision tower: CLIP-style ViT (hidden 1280, 32 layers, LayerNorm plus GELU, a pre-layernorm whose checkpoint key is spelled pre_layrnorm), patch_size 14, with native-resolution packing: image_grid_thw, cu_seqlens variable-length attention, 3D RoPE over (t, h, w) axes, spatial_merge_size 2, temporal_patch_size 2. mlxcel already implements this packing pattern in src/vision/qwen2_5_vl.rs and src/vision/qwen3_vl.rs; reuse the cu_seqlens and RoPE helpers.
• Projection: two stages, a per-patch multi_modal_projector (projection_dim 6144) followed by a patch_merge_mlp that folds spatial_merge_size^2 patches into the text hidden size. Visual tokens merge into the sequence by reshaping over the merge grid.
• Processor: dynamic resolution (smart resize to factor-aligned dims, patchify into (grid_t, grid_h/m, m, grid_w/m, m)), emitting pixel_values plus image_grid_thw; video is analogous. Config block img_token_compression_config.{spatial_merge_size, temporal_patch_size}.
• Token ids: image 200025, video 200026, vision start 200029, vision end 200030. Placeholder count per image is grid_t * (grid_h/2) * (grid_w/2).

Implementation plan

1. src/vision/minimax_m3_vl.rs (tower plus projectors) and a processor in src/vision/processors/; reuse the qwen2_vl helper functions where they match.
2. Loader arm in src/loading/ handling the checkpoint's actual dtype/layout split between quantized text weights and plain vision weights (verify against the real checkpoint, following the pattern of prior VLM ports).
3. Registration in vision detection, model_metadata, generate_vlm summary, TP arch string, and docs/supported-models.md.
4. Tests: grid math unit tests, placeholder count test, and a real-image smoke test.

Acceptance criteria

• A real MiniMax-M3-VL checkpoint answers image questions correctly through CLI and server.
• Placeholder expansion count matches the processor grid math for non-square images.
• Lints and format are clean.

Validation

Real checkpoint, at least two images with distinct contents (object identification) plus one multi-image prompt. If no public VL checkpoint is available yet, keep this issue blocked and note the availability check in a comment.

Effort: high. Blocked by #763.