feat(rl): add off-policy IS correction hook (current policy vs rollout)

[EazyReal]

What changed

• policy_loss_function (slime/backends/megatron_utils/loss.py) now passes the current grad-carrying log-probs as cur_log_probs into the TIS-correction kwargs, alongside the existing frozen train_log_probs (pi_theta_old) and rollout_log_probs (pi_rollout). The existing corrections (vanilla_tis_function, icepop_function) ignore the new kwarg via **kwargs.
• New off_policy_is_function in slime/utils/ppo_utils.py: a truncated-IS correction whose detached weight is clip(pi_theta / pi_rollout) against the actual rollout logprob, so one weight corrects both the train/inference mismatch and async (multi-version) staleness. The existing TIS only had pi_theta_old / pi_rollout, which equals this only in the single-update-per-rollout limit.
• Selected through the existing hook mechanism, no new arg or dispatch branch:
  --use-tis --custom-tis-function-path slime.utils.ppo_utils.off_policy_is_function.
• Clipping uses --eps-clip / --eps-clip-high (the CISPO/PPO clip convention); --eps-clip 1.0 gives canonical single-sided clipping. Note this reuses the policy-loss clip range rather than the --tis-clip* range used by the other TIS hooks.

Why

On a REINFORCE base (--advantage-estimator reinforce) this reproduces the CISPO surrogate (MiniMax-M1, arxiv 2506.13585) as a composable correction: L = -clip(pi_theta/pi_rollout).detach() * A * log pi, with gradient only through log pi. It generalizes the train/inference-mismatch correction to also absorb off-policy staleness without adding a dedicated estimator.

Validation

CPU unit test tests/test_off_policy_is.py (registered in the cpu-unittest matrix, NUM_GPUS = 0), run via pytest tests/test_off_policy_is.py:

• clamp behavior and clipfrac on known ratios, with loss_masks passed through unchanged;
• on a manual REINFORCE base, loss and gradient match the closed-form CISPO surrogate exactly (gradient flows only through log pi);
• --eps-clip 1.0 disables the lower bound (single-sided).
  The cur_log_probs wiring in loss.py imports megatron and is exercised by the GPU CI suites.

[EazyReal]

@zhuzilin could you review this one? It adds the off-policy IS hook at the policy-loss boundary using current-policy logprobs against rollout logprobs, preserving gradient flow through pi_theta while keeping the correction weight detached.