Refine training and conversion, support Qwen3.5 MoE and Gemma 4

[liang2kl]

• Adds support for quantizing and saving Qwen3.5 MoE models / Gemma 4.
• Fix changes in vllm 0.19.0 that break the per-shard Marlin conversion.

vLLM real quant will be implemented in a separate PR.

[liang2kl]

For now the MoE support is hard-coded for Qwen3.5. Will refactor this in the next PR.

[zhijian-liu]

Looks good overall. But we need to think about MoE integration complexity.