Add TurboT2AV SageAttention, SageSLA and FastNorm submodule

[liuyuxiang1021]

Summary

This PR adds TurboT2AV as an optional TurboDiffusion submodule for accelerated text-to-audio-video inference.

The submodule main branch is rebuilt from the original inference branch and only adds inference-time acceleration:

• SageAttention for selected unmasked LTX attention modules
• TurboDiffusion FastNorm for RMSNorm/LayerNorm paths, including the functional RMSNorm helper used by LTX
• TurboDiffusion SLA/SageSLA self-attention adapters with --sla_topk, --sla_block_q, and --sla_block_k controls
• inference CLI flags for acceleration and timing
• README timing notes for the 40-step teacher, accelerated 4-step student, and SageSLA speed/quality tradeoffs

Not included in this integration:

• training scripts or training workflow changes
• W8A8 quantization
• trained SLA adapter checkpoints

Submodule

• Repository: liuyuxiang1021/turbo-t2av
• Branch used by the submodule: main
• Current commit: c0e4506a40e9a4eb3ed5f34b0fe7dfcd3f26f084
• Clean base branch: origin_inference at c99f03b7b615661f63513b3816ea6c62b754c5ce

The parent repository .gitmodules tracks TurboT2AV on main, and the submodule gitlink is pinned to c0e4506.

Validation

Submodule checks:

• git diff --check
• pixi run python -m compileall packages/ltx-distillation/src/ltx_distillation/acceleration.py packages/ltx-distillation/src/ltx_distillation/tools/run_av_inference_eval.py
• CLI help includes --attention_type {default,sageattn,sla,sagesla}, --sla_topk, --sla_block_q, and --sla_block_k
• Secret/path scan found no real tokens, proxy settings, or local dataset paths in the submitted files

H20 generator-only measurements documented in the README:

• 512x768, 121 frames, 4 prompts: default 4-step student 2.53s/video; SageAttention self + FastNorm 2.17s/video (1.16x)
• 704x1280, 121 frames, first 4 prompts: SageSLA top-k sweep showed the quality/speed tradeoff; quality-first topk=1.0 reached about 1.10x with lower visual error than sparse settings
• 1024x1792, 121 frames, first 4 prompts: SageSLA topk=0.4 stress test improved median generator time from 15.97s/video to 10.58s/video (1.51x), with visible quality changes

The README also explains why the attention speedup is lower than TurboDiffusion Wan2.1 720p SageSLA results: the default LTX path already uses an efficient attention backend, TurboT2AV has a shorter latent sequence at normal resolution, and non-attention work remains unchanged.