Fix Diffusion Policy EMA with current Diffusers

[ruziniuuuuu]

Summary

• update Diffusers and Hugging Face Hub to a resolvable current cohort
• isolate the modern EMAModel API behind a private Diffusion Policy EMA owner
• preserve the existing checkpoint ema payload as averaged network weights
• restore the EMA tracker from loaded averaged weights for resumed training

Validation

• focused EMA update and checkpoint round-trip test
• clean Python 3.12 dependency resolution with uv pip compile setup.py

This addresses the same dependency update as #263 while also fixing its checkpoint serialization and restore path.