Add support for Flux and Wan models

[Abhinay1997]

This is a PoC showing how RAS works for Flux and Wan and not optimised at the kernel level.

Additional cmdline args have been added to support Wan 2.1

To run the Wan inference:

python wan_inference.py --height 480 --width 848 --is_video --num_frames 81 --seed 0 --sample_ratio 0.05 --prompt "A dog running on the beach" --negative_prompt "Low quality, blurry, messy, distorted" --output "out.mp4"

If you have any suggestions/fixes, I'll be happy to make changes.

Wan 2.1 1.3B samples:

Sampling ratio: 0.1
https://github.com/user-attachments/assets/af29b6b2-ad08-49a5-87f2-0d5dddacc0fd

Sampling ratio: 0.5
https://github.com/user-attachments/assets/7f5cefc6-d2db-4728-8d4d-13c0bed9de0c

Sampling ratio: 1.0 aka base model
https://github.com/user-attachments/assets/0c84987c-f121-4510-a086-0295fef08527

Flux samples:

Sampling ratio: 0.1

Sampling ratio: 0.5

Sampling ratio: 1.0 aka base model

[Abhinay1997]

Also made the change to support different height and width dimensions.

[Abhinay1997]

@microsoft-github-policy-service agree