test(grpo): add training loop tests with mocked environment

[abrichr]

No tests exist for the most complex parts of the GRPO training pipeline:

• _training_step
• _compute_rollout_loss
• _compute_ref_log_probs
• _make_agent_fn
• collect_group

Need mock-based tests that provide fake rollouts and verify:

• Gradient flow
• Loss computation correctness
• Edge cases (empty rollouts, zero advantages, all-same rewards)

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). TRL's GRPOTrainer is battle-tested with extensive upstream test coverage for training step, loss computation, and edge cases. We no longer maintain custom _training_step, _compute_rollout_loss, or _compute_ref_log_probs methods that need mocked-environment tests.