Add Strands SDK integration for RAG agent training

[JunjieAraoXiong]

Aim

Adding Strands SDK support to rLLM as an alternative to LangGraph for training RAG agents. Strands uses a simpler @tool decorator instead of LangGraph's graph-based setup.

Based on Tianhao's LangGraph example in examples/sdk/langgraph/.

Changes

New files in examples/strands/:

• retrieve_tool.py — retrieve tool that calls the RAG server
• run_strands.py — agent workflow + training interface
• train_strands_agent.py — training script (HotpotQA + RewardSearchFn)
• train_strands.sh — shell script for GPU training
• .env.example — env vars template

Bug fixes:

• rllm/integrations/strands.py — Fixed Qwen3 tool calling issue (see below)
• rllm/engine/rollout/openai_engine.py — Filter out Strands-specific kwargs

Qwen3 tool calling fix

When I tested on the GPU cluster, Qwen3-4B wasn't making proper tool calls. Instead of structured <tool_call> output, it was dumping JSON as plain text and rambling for 5-10 min per rollout.

Fixed with 3 changes:

1. Added fallback regex parser to catch tool calls in plain text
2. Set tool_choice="required" so model has to call a tool
3. Disabled Qwen3 thinking mode (enable_thinking=False)

Training setup

• RLOO algorithm, Qwen3-4B, 8x H100
• HotpotQA dataset, RewardSearchFn
• RAG server on port 9002

Status

Draft — waiting to verify LangGraph baseline first, then will run Strands training and compare the curves.

Testing done

• Local test with mock server works
• Trajectory saved correctly
• GPU training (pending)