[Model] Add MiniMax-H3 diffusion GGUF support - #119

Summary

• add a MiniMax-H3 diffusion adapter for GGUF DiT weights
• preserve non-DiT checkpoint names so vLLM-Omni model loaders retain ownership of model-specific mapping and fusion
• keep quantized fused QKV tensors intact while restoring dense QKV tensors to the checkpoint layout expected by MiniMax-H3
• enforce the generic one-gguf_model-to-one-weight-source invariant
• reject incompatible H3 schemas before model loading with an actionable error
• expose tested coverage at the family level as Video generation | MiniMax-H3 | Q4_K_M

Supported scope

• one task-specific MiniMax-H3 DiT per process
• non-pruned GGUF schema with complete time_embedder weights
• text encoder and VAEs remain on base Hugging Face weights
• combined FL2VA + Ref2VA loading is not supported by one gguf_model
• end-to-end validation in this PR covers FL2VA; it does not claim Ref2VA end-to-end validation

Validation

• python -m pytest -q tests/diffusion -m "not slow": 47 passed, 2 deselected
• focused MiniMax-H3 adapter and diffusion-loader tests: 27 passed
• Ruff and pre-commit checks passed
• full leejet/MiniMax-H3-GGUF/minimax_h3_fl2va-Q4_K_M.gguf strict load and generation succeeded on one NVIDIA B300
• generated video shape: (107, 256, 256, 3)
• generated audio shape: (1, 2, 142400)
• observed peak GPU memory: 81644 MiB

This is one functional end-to-end validation run, not a latency or memory benchmark.

Validation environment

Documentation

• vLLM-Omni usage documentation: vllm-project/vllm-omni#6512