feat: parallelize rlm_query_batched subcalls for faster execution

[dhananjai1729]

Summary

Addresses #135 — rlm_query_batched now runs independent child RLM subcalls in parallel using ThreadPoolExecutor, reducing wall time from N × T to ~T for N I/O-bound subcalls.

• Parallel execution via bounded thread pool for 2+ prompts in _rlm_query_batched
• New max_concurrent_subcalls parameter (default: 4) on both RLM and LocalREPL, propagated to child RLMs at each recursion level
• Thread-safe completion collection with a local lock; results and _pending_llm_calls maintain original prompt order for deterministic metadata
• Zero overhead for single prompts — skips ThreadPoolExecutor when len(prompts) <= 1
• Partial failure isolation — exceptions in one thread don't affect others

Files changed

File	Change
rlm/environments/local_repl.py	Rewrote _rlm_query_batched with ThreadPoolExecutor; added max_concurrent_subcalls param
rlm/core/rlm.py	Added max_concurrent_subcalls param; propagated to environment and child RLMs
tests/test_rlm_query.py	Added 6 new tests for parallel execution

Test plan

• All 263 existing tests pass (10 skipped due to missing API keys)
• New TestRlmQueryBatchedParallel test class with 6 tests:
  • test_batched_runs_in_parallel — verifies wall time is significantly less than sequential
  • test_batched_respects_max_concurrent — verifies concurrency cap via active thread tracking
  • test_batched_preserves_order_parallel — verifies result order matches input order with random delays
  • test_batched_parallel_partial_failure — verifies failures in one thread don't affect others
  • test_batched_pending_calls_ordered — verifies _pending_llm_calls are in prompt order
  • test_single_prompt_skips_thread_pool — verifies no thread pool overhead for single prompt

cc @alexzhang13 @alt-glitch @lakshyaag @rawwerks — would appreciate your review and feedback!

🤖 Generated with Claude Code