Keep Z-Image target timesteps and sigmas aligned

[zjn20030811]

Summary

• synchronize Z-Image sigmas whenever target timesteps are inserted
• preserve the scheduler invariant timestep = sigma * num_train_timesteps
• add a CPU regression using the 8-step student and 50-step teacher schedules used by trajectory imitation

TrajectoryImitationLoss asks the teacher schedule to contain each student timestep exactly. The scheduler previously replaced only the timestep value at the nearest teacher slot and left its sigma unchanged. The model was therefore conditioned on the student time while the flow integration still used the old teacher sigma.

For the real 8/50-step schedules, the old mismatch between timestep and sigma * 1000 reached 13.16. Updating both values keeps teacher trajectory states aligned with the student schedule.

Testing

• python -m pytest tests/test_flow_match.py -q (1 passed)
• ruff check tests/test_flow_match.py
• ruff format --check tests/test_flow_match.py