fix: restore args.ckpt_step after load_other_checkpoint

[keepkeen]

What

--ref-ckpt-step / --opd-teacher-ckpt-step permanently leak into args.ckpt_step, so later checkpoint loads in the same process resolve against the wrong iteration.

Root cause

load_other_checkpoint restores args.ckpt_step only when the saved value was non-None, instead of when an override was applied:

old_ckpt_step = None
if model_tag == "ref" and self.args.ref_ckpt_step is not None:
    old_ckpt_step = self.args.ckpt_step          # normally None
    self.args.ckpt_step = self.args.ref_ckpt_step
...
if old_ckpt_step is not None:                    # False -> never restored
    self.args.ckpt_step = old_ckpt_step

args.ckpt_step is a user override that defaults to None, so in the normal case the restore never fires and the ref/teacher step stays in args.ckpt_step for the rest of the process.

Consequence

get_load_checkpoint_path_by_args honors the leaked value for every subsequent load (model.py):

if getattr(args, "ckpt_step", None):
    iteration = args.ckpt_step

__init__ loads in the order ref → teacher → old_actor, so:

• with --keep-old-actor, the old_actor / rollout_actor weights load from iteration <ref_ckpt_step> of the actor checkpoint — silently wrong weights for the importance ratio if that iteration happens to exist, a crash if it doesn't;
• with OPD but no --opd-teacher-ckpt-step, the teacher load inherits the ref's step the same way.

Fix

Save ckpt_step in the same old_args tuple the function already uses for load / no_load_optim / no_load_rng / finetune, and restore unconditionally — identical pattern, one more field.

Verification

slime/backends/megatron_utils/actor.py needs a live Megatron to import, so there is no CPU unit test to attach (none of actor.py is unit-tested today). The defect is a three-line control-flow inversion verifiable by inspection: the override is applied under ref_ckpt_step is not None, the restore under old_ckpt_step is not None, and the two conditions differ exactly when args.ckpt_step started as None — its default. The GPU CI paths that pass --ref-load (e.g. test_qwen3_4B_ppo.py) don't set --ref-ckpt-step, which is why this never surfaced there.