Support exporting environment-specific metrics

[tyler-griggs]

It is useful to see environment-specific metrics during training runs to understand how the model is progressing. To do so, we should support SkyRL-Gym environments returning metrics that will be reported to wandb (or other logging/reporting tools).

TODOs

• add a metrics field of type Dict[str,Any] to BaseTextEnvStepOutput
• add the logic in SkyRLGymGenerator to aggregate these metrics across the trainingbatch and return them to the trainer for logging (under env/* on wandb)
• update one of the example environments to report metrics (to serve as an example)

[alex-dr]

I just came here to file this feature request :)

[tyler-griggs]

@mdawn65 is picking this up