feat(rl): add REINFORCE advantage estimator

[EazyReal]

What changed

Adds a reinforce option to --advantage-estimator. It reuses the GRPO group-normalized advantages and applies the plain additive surrogate -A * log pi_theta (new compute_reinforce_loss in slime/utils/ppo_utils.py) — no importance-sampling ratio, no clipping; gradient flows only through log_probs, and clipfrac is identically zero.

Wiring:

• compute_advantages_and_returns (slime/backends/megatron_utils/loss.py): reinforce routed through the existing GRPO returns path (get_grpo_returns).
• policy_loss_function (slime/backends/megatron_utils/loss.py): dispatches to compute_reinforce_loss for reinforce.
• Reward group-normalization (slime/ray/rollout.py): reinforce added to the mean-centering and optional std-normalization sets, identical to GRPO.
• --advantage-estimator choices/help (slime/utils/arguments.py).

Why

REINFORCE with a group baseline is a useful low-overhead estimator: same group normalization as GRPO but without the PPO clip/IS machinery. It is the on-policy base that off-policy importance-sampling corrections can layer on top of.

Validation

CPU unit test tests/test_reinforce.py (registered in the cpu-unittest matrix), run with pytest tests/test_reinforce.py:

• compute_reinforce_loss matches the closed form -A * log_probs and returns all-zero clipfrac.
• Backprop produces d/d log_probs = -A (gradient flows only through log_probs).

End-to-end: with the dispatch wired into compute_advantages_and_returns, --advantage-estimator reinforce runs through the GRPO returns + group-normalization path and no longer raises NotImplementedError.

[EazyReal]

@zhuzilin could you review this one? This adds the missing REINFORCE estimator as a narrow advantage path, keeping the estimator interface aligned with the existing GRPO/RLOO style instead of adding a separate training special case.