fix: allow fast inference without control signals

[primorLee]

Summary

generate_fast.py exposes --action_path as optional, but WanI2VFast.generate() unconditionally split c2ws_plucker_emb after the camera-preparation block. When no control path was provided, that local variable was never defined and fast inference failed before the first DiT call.

This change keeps camera chunks optional and passes dit_cond_dict=None for uncontrolled generation, which is already supported by WanModelFast.forward(). The controlled path retains the same tensor chunking and condition dictionary.

Validation

• python -m py_compile wan/image2video_fast.py
• Executed the real WanI2VFast.generate() control flow with CPU fake model/T5/VAE components:
  • pre-fix no-control path reproduces UnboundLocalError
  • post-fix no-control path completes with output shape (3, 9, 32, 64)
  • controlled outputs before and after the patch are exactly equal