[CharlelieLrt]

Description

This PR adds optional loss hooks for users who integrate FastGen into an external
training loop. Shared reductions support masking and weighting for sCM, MeanFlow,
DSM, VSD, and adversarial losses. DMD2 and f-distill callers can select a training
phase explicitly. For f-distill, users can save running ratio statistics in
their own checkpoints and choose the process group that combines those statistics:

losses, outputs = model.single_train_step(batch, iteration, phase="auxiliary")
h, bins = f_divergence_weights(
    scores, t, weighting_fn, bins=bins,
    logit_weight=score_weight, statistics_group=data_group,
)

Existing calls keep their defaults, and the external loop continues to own
optimizers and checkpoints. TRAINING_INTEGRATION.md describes the masking
contracts and extension points. PhysicsNeMo can then delegate shared loss math
to FastGen instead of maintaining private compatibility reductions.

Dependencies

• Depends on

  • This PR starts from FastGen main and has no prerequisite PR.

• Depends on this PR

  • PhysicsNeMo #2205 replaces local compatibility code with the proposed FastGen loss hooks. It depends on acceptance of these hooks.

• Independent PhysicsNeMo stack

  • PhysicsNeMo #2199 adds named diffusion inputs and outputs.
  • PhysicsNeMo #2200 adds inverse SNR and TrigFlow components.
  • PhysicsNeMo #2201 adds private adapters and public sCM and MeanFlow losses.
  • PhysicsNeMo #2202 extends the loss API to DMD2 and f-distill.
  • PhysicsNeMo #2203 adds checkpoint bundles for restarting distillation.
  • PhysicsNeMo #2204 adds the optional training strategy and the grouped changelog entry.

• Independent packaging proposal

  • FastGen #43 allows the NumPy 2 dependency used by PhysicsNeMo. It does not depend on these hooks.

If FastGen maintainers decline FastGen #44, we would close PhysicsNeMo #2205.
The preceding PhysicsNeMo stack would remain available.

Review process

Reviewers can compare this PR with its immediate parent. After a parent merges,
we update the remaining branches and their PR bases to keep each diff focused.