Fix MiniMax H3 GGUF model and encoder loading

[Gingertrout]

MiniMax H3 I2V fails before sampling: Qwen3-VL-32B GGUF loading omits the vision projector, so ComfyUI selects Mistral and raises TypeError: the JSON object must be str, bytes or bytearray, not NoneType. MiniMax diffusion GGUFs without architecture metadata also fail detection.

Load and map the matching Qwen3-VL projector, including temporal patch embeddings and DeepStack mergers, and preserve the native key layout used to detect MiniMax's 32B encoder. Add MiniMax diffusion architecture recognition. Keep Qwen2-VL loading and smaller Qwen3-VL text-only workflows working; report a missing 32B projector directly.

Validation:

• 11 loader regression cases passed, including actual GGUF files with and without architecture metadata, projector selection/mapping, and existing text-only behavior. Two existing ComfyUI MiniMax model tests also passed.
• The MiniMax loading fixes completed a neutral 256x256, five-frame, two-step I2V smoke test on an RTX 3090 24 GB: image/text conditioning, sampling, video decode, and audio decode. Peak PyTorch allocation was 21.8 GiB. Text encoding also passed through the restarted live ComfyUI server.
• Tested models: minimax_h3_fl2va_pruned-Q4_K.gguf, Qwen3VL-32B-Thinking-Q4_K_M.gguf, and its matching Q8_0 projector; ComfyUI 0.36.0 (a2f455c9), PyTorch 2.9.1+cu128. This verifies execution, not equivalence to the reference Instruct encoder's output quality.

Related: #473, #476, #478, #480. This provides a small combined alternative against main with regression coverage for both failures. Please consider it for an upcoming release.