Add prime-rl-based RLM training harness and update the local REPL implementation.

[alexzhang13]

Add a verifiers-compatible training harness for rlm.RLM at depth=1, designed to plug into prime-rl for end-to-end RL training. Rollouts run through a subprocess-isolated local REPL (no cloud sandboxes), with sub-LM calls proxied back to the trainer's inference server. Minor tweaks to rlm/core/rlm.py, rlm/utils/parsing.py, and rlm/utils/prompts.py so the local REPL behavior stays aligned with the training surface.

• training/src/rlm_train/ — env, rubric, sub-LM proxy, and subprocess REPL worker
• training/environments/oolong/ — example OOLONG long-context QA env
• training/configs/rlm-qwen3-30b-example.toml — example RL config
• Small alignment fixes to core REPL prompts/parsing so local and training paths match