[feature request] log generation data to help debugging

[chenhaiq]

Lots of people hit training failures after 100 steps especially in multi-turn agentic RL.
For example AgentR1/Agent-R1#30 (comment)

This kind of problem is very difficult to debug due to lacking tools.

The idea in this issue is to log input\output from LLM and tool calls into external tracking system such as wandb.

[SwordFaith]

There are many potential reasons which hard to identify by current tracking, some general useful metric we need to add are:

• batch solve all / solve nan batch/
• policy_loss logp in actor/
• tokenization mismatch cnt, ratio, avg per trajectory in rollout/
• tools & conversations behavior stats (latency, call times, error_type, code_len（tool_specific） etc.) (I have a draft, can optimize together)
• Seperate tool behavior log system, only need when check tool consistency (I have a draft, can optimize together)