Add FastWan (DMD) distillation method

[bghira]

Currently very VRAM-heavy due to lack of VSA and USP in SimpleTuner.

[Copilot]

Pull Request Overview

This PR adds FastWan DMD (Distribution Matching Distillation) as a new distillation method to SimpleTuner. DMD is a memory-intensive distillation technique that uses a generator (student) and fake score transformer (discriminator) to achieve high-quality distillation in fewer steps (typically 3-step).

Key changes:

• Implements DMD distillation infrastructure with generator and fake score transformer components
• Adds comprehensive DMD configuration options and training logic
• Updates documentation to include DMD usage examples and comparison with DCM

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
helpers/distillation/factory.py	Adds DMD enum value and factory method for creating DMD distillers
helpers/distillation/dmd/distiller.py	Complete DMD implementation with generator/discriminator training logic
documentation/distillation/WAN_DCM.md	Minor update to validation steps from 50 to 4
documentation/distillation/FASTWAN_DMD.md	Comprehensive DMD documentation with configuration examples

Comments suppressed due to low confidence (1)

helpers/distillation/dmd/distiller.py:322

• [nitpick] The function name '_pred_noise_to_pred_video' is misleading as it suggests video-specific processing, but the function performs generic noise-to-clean conversion that works for any latent type. Consider renaming to '_pred_noise_to_clean_latents' or '_convert_noise_prediction_to_clean'.

    def _pred_noise_to_pred_video(self, pred_noise, noise_input, timestep):