feat(loss): split surrogate (--policy-loss) and IS granularity (--is-level) out of --advantage-estimator

[EazyReal]

What

--advantage-estimator currently decides three independent things: the advantage (credit assignment), the policy-loss surrogate (cispo), and the IS-ratio granularity (gspo). compute_advantages_and_returns (loss.py:716) maps grpo/gspo/cispo to the same get_grpo_returns, which shows gspo/cispo are not advantage choices at all — they are re-read as the surrogate (loss.py:974) and the granularity (loss.py:960). The consequence is that valid combinations (e.g. CISPO bounding on a GAE advantage) cannot be expressed.

This PR makes the three axes independent:

• --policy-loss {ppo, cispo} (default ppo) — the surrogate / bounding rule.
• --is-level {token, sequence} (default token) — the IS-ratio granularity; sequence is GSPO.
• --advantage-estimator now means credit assignment only.

Behaviour preservation

Dispatch-only — compute_policy_loss / compute_cispo_loss / compute_gspo_kl are untouched, so every existing preset reproduces identically. Legacy --advantage-estimator values are remapped for one release with a FutureWarning:

legacy	maps to
--advantage-estimator gspo	--advantage-estimator grpo --is-level sequence
--advantage-estimator cispo	--advantage-estimator grpo --policy-loss cispo

The legacy values are kept in the --advantage-estimator choices (so old commands still parse) and remapped in slime_validate_args; a follow-up can drop them from the choices once the deprecation window closes.

Changes

• slime/utils/arguments.py — add --policy-loss / --is-level; deprecation shim in slime_validate_args; the CISPO single-sided-clip warning now keys off --policy-loss.
• slime/backends/megatron_utils/loss.py — surrogate dispatch keys on --policy-loss, granularity (incl. need_full_log_probs) on --is-level; drop the now-unreachable gspo/cispo literals from the advantage branch.
• slime/backends/megatron_utils/actor.py — the log-prob-reuse optimisation gate keys on --is-level instead of == "gspo".
• slime/ray/rollout.py — drop the dead gspo/cispo literals from the reward-normalisation groups (legacy values are remapped to grpo before this runs).
• tests/test_megatron_argument_validation.py — unit tests for the shim (gspo/cispo remap + warning) and for orthogonal pass-through; the validate-args factory gains the new fields.

Testing

tests/test_megatron_argument_validation.py — 19 passed (16 existing + 3 new) under a CPU venv. No GPU required.

Scope / non-goals

• No change to loss numerics, the Megatron/CP machinery, the reducer, or value/SFT losses.
• REINFORCE (feat(rl): add REINFORCE advantage estimator THUDM/slime#2083) slots in as a --policy-loss reinforce value once it rebases onto this.
• Decoupled reference structure, a surrogate registry, and additive regularizer assembly are separate follow-ups.

Implements increment #1 of the loss-axis RFC (#1).