feat: adapt Agent-R1 to verl 0.8.0

[lqzxt]

Summary

Adapt Agent-R1 to verl==0.8.0 by migrating training, reward, rollout, and worker integration to the updated verl APIs.

Changes

• Switch actor/critic execution to the unified model engine worker path and remove legacy FSDP/DataParallel actor/critic overrides.
• Update PPO trainer logic to use tensor-dict mini-batch updates, extract_reward, and the new reward config namespace.
• Replace legacy agent loop server manager usage with LLMServerClient and GlobalRequestLoadBalancer.
• Update reward loop imports and reward model config paths for verl 0.8.0.
• Ensure rollout weight sync calls update_weights(mode="naive") before waking vLLM replicas.
• Update recipe agent flow type hints/imports for the new LLM client.
• Add a GSM8K GRPO launch script.
• Update README and installation docs from verl==0.7.0 to verl==0.8.0.

[gemini-code-assist]

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.