Source sync: Add Flax configuration updates and improve WAN trainer profiling

[michelle-yooh]

This PR contains source synchronization from internal MaxDiffusion updates with the following improvements:

Changes:

Flax Configuration - Add automatic Flax config update to disable variable sharding by default in the main init.py, ensuring consistent behavior across the codebase.

Configuration Loading Fix - Improve model configuration loading in configuration_utils.py to prioritize subfolder-specific configs when available. This ensures models with subfolder structures load their configs correctly before falling back to the root directory.

WAN Trainer Profiling Enhancements - Add jax.named_scope() annotations in wan_trainer.py to improve performance profiling:

• Wrap the forward pass in a named scope for better profiling visibility
• Wrap loss computation in a named scope to separately track loss calculation overhead
• Enables clearer performance metrics and debugging in XLA compilation

These changes improve model loading robustness, provide better profiling granularity for performance analysis, and ensure consistent Flax behavior across the framework.

PiperOrigin-RevId: 866156279

[github-actions]

e2e testgrid: https://8bcf50593faf4ea38060e236169827e5-dot-us-central1.composer.googleusercontent.com/dags/maxdiffusion_tpu_e2e/grid