[examples, docs] feat: add end-to-end Agent RL training recipes

[yyDing1]

What does this PR do?

This PR adds end-to-end Agent RL training recipes for both white-box ReAct and black-box Claude Code Agents using verl's V1 colocate_async trainer.

Key changes

• Add runnable training recipes for:
  • Qwen3-Coder-30B-A3B-Instruct with ReAct.
  • Qwen3.5-4B with Claude Code.
• Add Task Configs for SWE-reBench training and SWE-Bench Verified validation.
• Document data preparation, Ray Runtime Environment, distributed launch parameters, shared storage layout, and output monitoring.
• Expose RLOO/GRPO, DPPO, clipping, rollout correction, and Router Replay settings through environment variables.
• Add per-runner trajectory_selection with backward-compatible all and new longest policies.
• Configure Claude Code training to keep the trajectory with the most model-generated tokens before reward assignment and TransferQueue writes.
• Improve sandboxed Claude Code invocation with explicit model and permission-mode configuration.
• Update the legacy MoE recipe to use separated data/runtime paths and Ray task dispatch.
• Replace example runtime credentials with placeholders.

Compatibility

Existing runners retain trajectory_selection=all by default. ReAct behavior is unchanged, while the Claude Code recipe explicitly enables longest.

Validation

• Framework trajectory tests: 20 passed.
• Added coverage for trajectory selection and invalid configuration.
• Updated Claude Code argument and environment tests.
• Validated training script syntax and Python compilation.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.