[RL] Improve Prime verifier integration

[taivu1998]

Add the env prepare lifecycle hook and an OpenAI-compatible verifier client, rewrite PrimeIntellectEnv for supported single-turn chat environments, and thread response loss masks through rollout batching. This makes Prime verifier training usable while preserving existing single-turn RL behavior.

[taivu1998]

🤖 Problem: PrimeIntellectEnv was not reliable for Marin RL because setup happened inside sampling, verifier generation only cleanly matched one backend path, Prime rollouts used a custom result processor, and Marin had no first-class way to mask response tokens for future multi-turn verifier trajectories.

Approach: Add a generic environment prepare lifecycle hook, add an OpenAI-compatible verifier client path for inference contexts, rewrite PrimeIntellectEnv around an explicit Phase 1 support matrix for single-turn chat verifier environments, remove the Prime-only vLLM processor, and add response_loss_mask plumbing from Rollout through training-batch construction.

Key code: Prime setup and support gates live in lib/marin/src/marin/rl/environments/prime_intellect_env.py, the verifier client contract lives in lib/marin/src/marin/rl/environments/inference_ctx/base.py and lib/marin/src/marin/rl/environments/inference_ctx/openai_compat.py, and the generic loss-mask threading lives in lib/marin/src/marin/rl/types.py and lib/marin/src/marin/rl/train_batch.py.

Tests: Verified with ./infra/pre-commit.py --all-files --fix, UV_CACHE_DIR=/tmp/uv-cache-pr-push uv run pytest tests/rl -m 'not slow' -q, and focused RL slices covering PrimeIntellectEnv, inference contexts, train-batch masking, replay/storage, and rollout worker behavior.