Added support for WAN 2.2 model

[prishajain1]

Added support for WAN 2.2 model

This PR introduces support for the WAN 2.2 model and refactors the existing pipeline to dynamically handle multiple model versions.

New Model Support (WAN 2.2)

The configuration for the WAN 2.2 model are added:

base_wan_27b.yml: New configuration file for the WAN 2.2 (27B) model.

Refactoring for Multi-Model Support

To allow the system to select between WAN 2.1 and 2.2, several core files were modified:

generate_wan.py: calls the correct pipeline (WAN 2.1 / WAN 2.2) depending on the model_name parameter in config files.
wan_utils.py: Updated all transformer loading functions to include a subfolder parameter.
base_wan_14b.yml: Added the model_name parameter to conform to the new configuration standard.
wan_pipeline.py: The processing pipeline which supports both WAN 2.1 and WAN 2.2 by using separate base classes
wan_checkpointer.py: Handles checkpoint loading/saving for both WAN 2.1 and WAN 2.2.
wan_checkpointer_test.py: Tests for WAN 2.1 and WAN 2.2 checkpointer

We tested the above pipeline for both WAN 2.1 and WAN 2.2 and the benchmarking numbers can be found here: https://docs.google.com/spreadsheets/d/1GJqOoTUikAqLWedQG0Qu4YQbdQ5IeveHMqnOgi3d1tY/edit?pli=1&resourcekey=0-ywo0BVMZ-W1Vf7kCJhekXQ&gid=1962169097#gid=1962169097

[entrpn]

Can you also make sure Wan2.1 training and inference still runs without issues before this PR goes into main?

[prishajain1]

Can you also make sure Wan2.1 training and inference still runs without issues before this PR goes into main?

sample logs inference:
WAN 2.1: https://paste.googleplex.com/6392513959165952
WAN 2.2: https://paste.googleplex.com/4651918836236288

sample logs training:
https://paste.googleplex.com/6353248971784192
https://paste.googleplex.com/6606675631013888

[entrpn]

https://paste.googleplex.com/6353248971784192

Is there a reason why training logs don't show training steps?

[prishajain1]

https://paste.googleplex.com/6353248971784192

Is there a reason why training logs don't show training steps?

By default the log_period = 100, we were running this for a shorter number of training steps.
Logs with log_period set to 1 : https://paste.googleplex.com/5210776497422336