E9 effective non dataset non actionnorm split

[Teeeio]

PR Category

Train

PR Types

Bug Fixes

PR Description

This PR keeps only the verified effective fixes on top of e9c9ab983881f7ce5d4f9439754884422c18aed5,
excluding dataset-mixing and action-normalization changes.

Included changes:

• Restore policy.train() in training loop to ensure training mode is enabled.
• Make visual input order deterministic via image_key_order (default: image -> wrist_image).
• Add robust image key resolution in Qwen VLM input path:
  • data.vla_data.image_features
  • fallback to data.image_key_order
  • fallback to runtime batch keys
• Read video_backend from config (with safe fallback).
• Read num_workers from config instead of hard-coding.
• Add concise code comments aligned with existing style.

Explicitly NOT included:

• Multi-dataset mixing (data_paths, mixture_mode, StarVLAMixtureDataset)
• Action norm dim changes (normalize_action_dims and related pre/postprocessor wiring)