Support PTQ for Qwen3 Omni

[tommyip]

Detailed description of the requested feature

Qwen3 Omni is a relatively new omni model that can process text/audio/image inputs and text/audio outputs. Would be nice if we can quantize it using modelopt to NVFP4 or AWQ.

The model consists of a few submodules:

• Vision encoder (Same as Qwen3 VL Vision Encoder)
• Audio encoder (Stack of conv2d + stack of encoder-decoders)
• Thinker (Qwen3 30B A3B)
• Audio output: Talker (Qwen3 MoE scaled down with some minor variations)
• Audio output: Multi-token prediction (Qwen3 MoE scaled down with some minor variations)
• Audio output: Code2Wav (ConvNet)

Describe alternatives you've considered

Qwen3 Omni is supposed to be supported by GPTQModel but has vram memory leak issue.

Target hardware/use case

Blackwell

[Edwardf0t1]

Thanks @tommyip for bringing this to us.

cc @omrialmog who can help on the prioritization of the Qwen3 Omni support.