Qwen3 omni moe support

[LRL2-ModelCloud]

Completes #1848

[Qubitium]

Refractor: Add new BaseQModel properties

ATTENTION_MASKS_DTYPE = torch.bool # default to bool
ATTENTION_MASKS_REQUIRED_FOR_INPUT : bool = False  # default to false

[Qubitium]

@LRL2-ModelCloud

 def pre_quantize_generate_hook_end(self):
        self.model.thinker.model.embed_tokens = self.model.thinker.model.embed_tokens.to(CPU)
        self.model.thinker.visual = self.model.thinker.visual.to(CPU)
        self.model.thinker.audio_tower = self.model.thinker.audio_tower.to(CPU)

        self.model.thinker.visual.rotary_pos_emb = self.model.thinker.visual.rotary_pos_emb.to(CPU)
        self.model.thinker.model.rotary_emb = self.model.thinker.model.rotary_emb.to(CPU)

Actually, call offload_to_disk() on these 3 modules so they directly go to disk and not waste cpu mem.

[Qubitium]

@LRL2-ModelCloud

add

INPUT_EMBEDDING_EXTRA_ARGS = None

This is for models like qwen vl whre we need to pass return_audio=False or return_video=False like args. We pass it as **INPUT_EMBEDDING_EXTRA_ARGS if not None.