Implement multi-environment and multi-domain mixing

[rafapi]

Implement multi-environment + multi-domain + multi-dataset training support:

• dispatcher/world/actor now accept per-domain overrides
• new coding and fn_calling domains (dataset loaders, rollouts, verifiers) plug into the registry
• configs/domain-mix configs were updated so mixed experiments can pull the required datasets (with proportional mixing), environments, and reward logic without manual wiring.