Feat: Local Deployment, Distributed Acceleration, Parallel VAE Decode, QKV Fused_projection, and torch.compile Support

[Sallyeen]

This PR introduces several major enhancements to the FramePack:

• Local Service Deployment: Added support for running and invoking models as local services.
• Distributed Inference Acceleration: Enabled parallelism with Ulysses and Ring Attention for faster distributed inference.
• Parallel VAE Decode: Optimized VAE decoding by introducing parallel execution.
• QKV Projection Fusion: Improved efficiency with optional fused QKV projection operations.
• Torch.compile Integration: Leveraged torch.compile to further optimize runtime performance.

[Sallyeen]

Performance Improvement

• Before optimization: Single RTX 4090, deployment time ≈ 237s
• After optimization: 8× RTX 4090, deployment time ≈ 49s
• Speedup: ~4.8× faster

Notes

These improvements have been tested in both local and distributed environments, showing significant gains in deployment efficiency and inference scalability.

Finally, I would like to sincerely thank you and the community for your outstanding contribution to this project.
It has provided an excellent foundation for further optimization, and I hope this PR can also be helpful in return.
Looking forward to your feedback!