Add ability to give custom rollout log function | Custom log hook where users can define how they want to log samples.

[yeshsurya]

Feature request

• New config: trainer.custom_rollout_log_function — a dotted Python path
• Hook receives keyword args: step, inputs, outputs, scores, ground_truths, reward_extra_infos, batch (raw DataProto), timing, config
• Return True → skip default logging; False → default continues
• Called for both train and eval rollouts
• Zero overhead when not configured
• Includes an example showing wandb Table logging of per-sample prompt/response/reward
• Similar to slime's --custom-rollout-log-function-path

Motivation

So that I can use different metrics providers like wandb etc to see how my rollouts are looking

Your contribution

I'll try to contribute this feature.