[codex] Fix Qwen3.5 MRoPE VLM model

[eligotts]

Summary

Fixes the custom Qwen3.5 MoE VLM path to match the MRoPE behavior expected by the upstream/HF-vLLM-style implementation.

• Adds a Qwen3.5 MRoPE position builder that uses renderer-provided mm_token_type_ids plus image_grid_thw instead of trainer 1D positions.
• Replaces the generic rotary embedding path with Qwen3.5-compatible interleaved MRoPE handling.
• Makes VLM forward compute image MRoPE internally, reject invalid 2D image positions, validate image token/feature counts, and accept explicit packed seq_lens for future packed multimodal batches.
• Adds targeted unit coverage for text-only positions, single-image positions, packed segment resets, packed-vs-standalone equivalence, video rejection, and HF rotary parity.

Validation

• FLA_TILELANG=0 uv run pytest tests/unit/utils/test_sequence.py tests/unit/train/models/test_qwen3_5_moe_mrope.py tests/unit/train/models/test_qwen3_5_moe_vlm.py tests/unit/train/models/test_qwen3_5_moe.py
• Also run as part of the combined focused suite: FLA_TILELANG=0 uv run pytest tests/unit/utils/test_sequence.py tests/unit/orchestrator/test_batch.py tests/unit/train/test_model_forward.py tests/unit/train/models/test_qwen3_5_moe_mrope.py tests/unit/train/models/test_qwen3_5_moe_vlm.py tests/unit/train/models/test_qwen3_5_moe.py