feat(models): add Cohere Compass quantization support

[ZX-ModelCloud]

Summary

• add a cohere_compass model definition for CohereLabs/North-Micro-Vision-Instruct image-text calibration and text-decoder quantization
• preserve layer-specific sliding/full attention masks, MRoPE embeddings, and DeepStack visual residuals during isolated GPTQ and AWQ layer replay
• make AWQ preservation of adapter-supplied position embeddings an explicit model opt-in, retaining the existing rebuild behavior for other models
• extend AWQ handling for Cohere Compass layer normalization

Validation

• pytest -q tests/models/test_cohere_compass.py tests/test_awq_rotary_device.py -k 'not TestCohereCompass' (14 passed, 1 deselected with the integration model available)
• pytest -q tests/test_gemma4_support.py tests/test_gemma4_unified_support.py (12 passed)
• focused Ruff checks for the changed adapter, AWQ, and test files
• git diff --check
• GPTQ and AWQ save/reload and image-generation smoke coverage
• the FAST ARC Challenge/GSM8K baselines use recorded scores