fix(train): skip optimizer and scheduler for eval-only

[EazyReal]

Replaces #2109.

Desired vs bug

Eval-only is intended. train.py already has:

# special case for eval-only
if args.num_rollout == 0 and args.eval_interval is not None:
    ray.get(rollout_manager.eval.remote(rollout_id=0))

The train loop is range(..., num_rollout), so this correctly does zero training steps. Eval still needs the actor loaded and update_weights() into the rollout engines, so create_training_models runs first.

The bug is that bring-up always constructed Megatron's optimizer and OptimizerParamScheduler. With --num-rollout 0 and default --lr-decay-iters, slime sets train_iters = 0 → lr_decay_steps = 0, and Megatron asserts lr_decay_steps > 0. You never reach the eval special case.

Why not train_iters = 1

That was #2109. It lies to the scheduler ("there is one training iteration") so the constructor succeeds, then still runs zero train steps. train_iters stops meaning "how many steps we train." It also still dies if lr_warmup_steps >= lr_decay_steps.

Eval-only does not train, so it should not build a train stack.

Change

• setup_model_and_optimizer: after get_model, if num_rollout == 0, set no_load_optim and return (model, None, None). Weight load is the same path load_other_checkpoint already uses for ref/teacher (optimizer=None, no_load_optim=True).
• create_training_models: do not allocate a critic when num_rollout == 0.

num_rollout > 0 is unchanged.

Test plan

• --num-rollout 0 --eval-interval 1 loads the actor, skips optimizer/scheduler construction, hits rollout_manager.eval
• A normal training run still builds optimizer + scheduler