Use native flow prediction for DMD critic loss

[zjn20030811]

Summary

• use the fake critic's native flow prediction for flow-matching loss
• avoid a flow-to-x0-to-flow round trip through two different scheduler grids
• preserve the existing non-flow prediction path

Why

WanDiffusionWrapper already returns both the model's flow prediction and its derived x0 prediction. The fake critic uses the wrapper's default shift of 8, while the generator scheduler in the default DMD configurations uses shift 5. Reconstructing flow from x0 with the generator scheduler therefore selects sigma from a different discretized grid and is not an identity operation.

For example, at the valid critic timestep 20 the old round trip scales the native flow prediction by about 1.194. Passing the native prediction directly to FlowPredLoss keeps it in the parameterization produced by the critic and matches the noise - x training target.

This addresses the mismatch discussed in #35.

Tests

• verifies that flow critic loss receives the native flow tensor with the expected flattened batch/frame layout
• verifies gradients propagate to the native flow prediction
• verifies the non-flow branch still converts x0 to a noise prediction

2 passed

[zjn20030811]

Follow-up in 1d1c610: the repository has a parallel long_video/model/dmd.py critic path that still reconstructed flow from the derived x0 prediction. It now consumes the native flow tensor returned by fake_score, matching the main DMD path and preserving the correct parameterization/gradient. Added a focused long-video regression case; python -m pytest -q tests/test_dmd_flow_prediction.py passes (3 tests).