Fix ignored seed and sigma shift during sampling

[primorLee]

Summary

The public MOVA inference API accepts both seed and sigma_shift, and the official inference scripts expose them, but neither value currently reaches sampling:

• prepare_latents and prepare_audio_latents are called with generator=None, so seed does not control the pipeline RNG.
• scheduler.set_timesteps is called without shift=sigma_shift, so non-default sigma shifts are ignored.

This change creates one CPU generator from seed and passes it through the video and audio noise initializers, then forwards sigma_shift to the scheduler. Reusing one generator keeps the complete joint video-audio sample deterministic, while seed=None preserves global-RNG behavior.

The default output schedule is unchanged for the released checkpoints because their scheduler configuration and the public argument both default to shift 5.

Validation

• Mocked the full pipeline setup through latent initialization and verified that the same seed reproduces both video and audio noise after perturbing the global RNG.
• Verified that changing the seed changes both modalities.
• Verified that sigma_shift=4.25 is forwarded to set_timesteps together with the requested step count and execution device.
• Ran Python 3.12 syntax compilation on Windows and in a clean Linux container.
• Ran git diff --check.