Make action-policy objectives padding-aware

[zjn20030811]

What changed

• apply the same ImageNet normalization to DiffusionPolicy inputs that the ACT and CNNMLP policies already use;
• compute ACT L1 and diffusion noise-prediction losses over valid action elements only, so episode-boundary padding cannot change the gradient scale;
• keep the all-padding edge case finite and add focused regression tests.

Why

The diffusion branch received the loader's [0, 1] images without the feature normalization used by the other policies. Its masked loss then called mean() over the padded tensor, so two otherwise identical batches with different amounts of padding optimized different effective objectives. The shared masked reduction makes the objective invariant to chunk padding while preserving the standard mean when every timestep is valid.

This is separate from the existing open questions about dependency setup and does not duplicate an open pull request. It also resolves the behavior described in #48.

Validation

• python -m pytest -q tests/test_losses.py (4 passed)
• python -m py_compile policy.py losses.py
• git diff --check