Qwen3.5-MoE on CUDA: load ExLlamaV3 (EXL3) checkpoints #1149

Native CUDA Qwen3.5-MoE has no loader for ExLlamaV3 (EXL3) checkpoints. I would like to add one, as two small pull requests, and I am opening this first because CONTRIBUTING.md asks for an issue before a new checkpoint format.

What: a Hugging Face-layout Qwen3.5-MoE checkpoint whose config.json says quantization_config.quant_method: "exl3" (turboderp's ExLlamaV3 trellis format, mul1 codebook, per-tensor widths) would load and serve on the existing cuda.exl3 Dense and grouped-expert kernels, with bf16 embeddings. The test checkpoint is https://huggingface.co/Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw (Ornith-1.5-122B, about 4 bits per weight, 60 GiB on the device). A second, independent change lets the Qwen vision tower read Hugging Face tensor names (model.visual.*).

Why: that checkpoint serves from one GB10 (121 GiB unified memory), where the MLX 4-bit form of the same model is larger.

Scope and limits: serving only, with the generic EXL3 kernels. There are no tuned prompt or decode kernels and no drafter (--drafter is refused for an EXL3 checkpoint), so speed is modest. The MTP layer and drafting are left out on purpose.

Related open work, not duplicated here: #1137 (dense BF16 Qwen3.5 scoring, touches the same geometry, engine and a cuda_bf16 file, so my files are named *_bf16_rows), #1145 (ModelOpt NVFP4 for Qwen3.5), #1113 (EXL3 for Flash Next), #1125 (Ornith on Metal), #1047 (padding rows in the routed experts).

Credit: the format and kernel design are turboderp's ExLlamaV3 (https://github.com/turboderp-org/exllamav3).