Add deep recursion support and vLLM docker quickstart

[Zhi0467]

Summary

• add deep recursion support with depth-specific backends and recursive client spawning, see rlm.clients.recursive
• update examples.quickstart to exhibit depth-2 recursion
• add logging for max recursion depth reached and number of llm_query calls per depth
• update docs/examples for recursion and vLLM usage
• add vLLM+Docker quickstart example and depth call aggregation fix
• rename max_depth to recursive_max_depth in RLM API for clarify, that arg in the initial RLM object controls max depth of the recursion tree

Note on design choice

• the max turns decay by a factor of 2 as we move down the recursion tree to prevent sub-RLMs from exploding
• the llm_query timeout is bumped to default 900, and decay by subtracting 120 with every one depth deeper

Testing

• uv run pytest
• uv run python -m examples.quickstart