Ensure tool calls and reserved names aren't overwritten + add logger metadata

[alexzhang13]

1. Scaffold / namespace protection. From #95

• Problem: Executed code could overwrite context, llm_query, FINAL_VAR, etc., and break RLM.
• Fix: Single set RESERVED_TOOL_NAMES in base_env (includes history). After each run:
  • LocalREPL: _restore_scaffold() restores globals and context/history from context_0/history_0.
  • Isolated envs (modal, prime, e2b, daytona, docker): in exec script, restore context and history from context_0/history_0 before save_state(_locals).
• Tests: TestLocalREPLScaffoldRestoration in test_local_repl.py.

2. Trajectory metadata on RLMChatCompletion

• RLMChatCompletion.metadata: Optional dict | None (default None) holding full trajectory (run_metadata + iterations with code blocks and sub-calls) so runs can be reconstructed.
• RLMLogger:
  • Always captures in memory. log_dir=None → only in-memory (for completion.metadata). log_dir="path" → same + write JSONL to disk.
  • clear_iterations() at start of each completion; get_trajectory() returns {run_metadata, iterations} for the current completion.
• RLM: Calls logger.clear_iterations() at start of completion(); sets metadata=logger.get_trajectory() on returned RLMChatCompletion.
• README: short “Trajectory metadata and logging” section (in-memory vs disk).
• Tests: TestRLMChatCompletion in test_types.py (metadata default + roundtrip).