feat(models): port GLM-OCR (ViT + GLM-4 text OCR)

[inureyes]

Summary

Adds GLM-OCR (glm_ocr, GlmOcrForConditionalGeneration) support, a document-OCR sibling of the in-tree GLM-4V stack. A 24-block ViT vision tower feeds a 16-layer GLM-4 text decoder driven by full-width even/odd MRoPE. Closes #547.

Architecture

• Vision tower (src/vision/encoders/glm_ocr.rs), a variant of the glm4v encoder:
  • Per-head q/k RMSNorm over head_dim (64) before the 2D vision rotary (glm4v has none).
  • Block norms take rms_norm_eps (1e-5); no learned position embedding and no post-conv norm (the checkpoint ships neither).
  • patch_embed and downsample detect both the channels-last export layout and the raw channels-second checkpoint layout.
  • Patch-order fix (OCR-critical): patches are reordered from the processor's raster order into spatial-merge-window order after the patch embedding, so the rotary, the consecutive-4 downsample, and the text-side merged-token grid all agree. glm4v leaves this latent misalignment in place (fine for coarse VQA, but OCR reads scrambled patches wrong).
• Text backbone reuses Glm4vTextModel unchanged. The loader lifts the nested rope_parameters block (mrope_section [16, 24, 24], partial_rotary_factor 1.0) into the fields the config deserializes, and drops the next-n prediction (MTP) layer at index num_hidden_layers.
• Processor reuses Qwen2VLProcessor with the OCR pixel bounds (min 12544, max 9633792).
• Wiring: detection -> loader -> LoadedModel::GlmOcr -> VlmRuntimeRef::Qwen -> server/CLI, plus model_metadata and TP dispatch rows.

Validation

Validated on mlx-community/GLM-OCR-4bit through the actual CLI runtime:

• Text Recognition on a spatially-asymmetric 4-quadrant image returns ALPHA 111 BRAVO 222 / CHARLIE 333 DELTA 444 in correct spatial reading order (clean text, no character scrambling), confirming the patch-order fix.
• Table Recognition emits correct structured HTML matching every cell in row/column order.

Automated tests:

• Co-located unit tests: conv-layout detection (both patch-embed and downsample layouts), the raster->merge-window permutation, rope_parameters lifting, and the MTP-layer drop.
• tests/glm_ocr_parity.rs: config-detection tests (run in CI without a checkpoint) plus checkpoint-gated real-model load and text-forward tests.

cargo test, cargo fmt --all -- --check, and cargo clippy --all-targets --features metal,accelerate are clean for the new code.