Dedupe echoed user text in predictions

Map run id to deployed checkpoint and add reward distributions

Add grouped result plots to technical report

Fix checkpoint id for Interactive-RL default

Use final LoRA checkpoint in RL default

Initial poor mans interaction models demo