fix(args): reject --kl-coef for GRPO-style estimators

[YeonwooSung]

Summary

--kl-coef is reward-KL shaping. GRPO-style estimators (grpo, gspo, cispo) do not apply KL to rewards, so a non-zero --kl-coef is a silent no-op today.

This follows the maintainer decision on #399: we should not use KL in the GRPO reward, and should use kl_loss instead. #1247, which applied KL to GRPO rewards, was closed unmerged. This PR does not change get_grpo_returns() or invent reward-KL math.

slime_validate_args now raises a clear ValueError when --kl-coef != 0 is combined with --advantage-estimator {grpo,gspo,cispo}, and points users at --use-kl-loss / --kl-loss-coef.

Test plan

• python -m pytest tests/test_megatron_argument_validation.py -q --tb=short
  • grpo / gspo / cispo + --kl-coef 0.1 raise ValueError matching kl-coef / kl-loss
  • grpo + --kl-coef 0 still passes
  • existing validate-args tests still pass

Fixes #397