Fix correlated video-action inference noise

[zjn20030811]

Summary

• draw video and action initialization noise from disjoint portions of one seeded random stream
• keep seeded joint inference exactly reproducible
• preserve the existing single-shape noise helper behavior

Why

Joint inference currently creates a new generator with the same seed for each modality. With the default Wan2.2 shapes, all 1,536 action-noise values are identical to the prefix of the video-noise stream (1,536/1,536). Training uses separate sequential randn_like draws, so inference starts from a different joint noise distribution.

Using one generator for both draws restores non-overlapping samples while retaining deterministic replay. In the focused default-shape check, prefix identity drops from 1,536/1,536 to 1/1,536; rerunning with the same seed reproduces both tensors exactly.

Testing

• python -m pytest -q -p no:cacheprovider tests/test_noise_utils.py
• Ruff checks on the new helper and test
• Python bytecode compilation for the changed modules and test