Custom metrics in reward function

**Is your feature request related to a problem? Please describe.**
Currently, the reward score is the only metric logged with the logger. The ability to report custom metrics (as well as standard min/max/mean breakdown) would be great.

**Describe the solution you'd like**
The ability for reward functions to return something more than reward, correct. e.g. a dict with `{"reward": ..., "additional_metrics": {...}}` or similar. I believe verl does something similar, expecting `score` (as reward) then additional metrics as all additional kwargs.

**Additional context**
An example here would be if the reward is a composite of multiple metrics, reporting each.