Reproduce Jan models with Verl codebase

[nguyenhoangthuan99]

We need to learn Verl codebase and try to reproduce the result of Jan-v1, Jan-nano to ensure consistency.

• Tools
• multi turn Agent loop
• roll out
• custom reward with chat history
• wandb logger for custom reward
  Verl Documentation: https://verl.readthedocs.io/en/latest/sglang_multiturn/multiturn.html#multi-turn-rollout-support

[nguyenhoangthuan99]

Current requirement is train with LoRA to reduce memory consumptions, but for verl,

• currently, they haven't supported multiturns + tools training env with vLLM rollout yet GRPO Multi-Turn Training with LoRA + vLLM volcengine/verl#3038.
• SGLang also not supported LoRA training GRPO LoRA training with FSDP + SGLang volcengine/verl#2110

-> try to reproduce Jan-v1 model results with SGLang rollout first, without LoRA training
-> implement our own multiturn + tool training env with vLLM rollout and contribute to upstream