fix: skip DDP and optimizer for prm_teacher (inference-only mode)

[HeJW0303]

Summary

PRM Teacher only performs forward passes to compute teacher log-probabilities for OPD loss. It never does backward passes or gradient updates. However, the current code loads it with full DDP
wrapping + Adam optimizer, wasting ~160GB of memory and causing OOM on single-GPU deployments.

Changes

File: slime/slime/backends/megatron_utils/model.py

1. New function _setup_inference_only_model(args, role): Creates model with wrap_with_ddp=False, avoiding DDP gradient buffer allocation.
2. New branch in initialize_model_and_optimizer(): When role == "prm_teacher", calls the inference-only setup, loads checkpoint without optimizer/scheduler, and returns early.

Memory Impact (27B BF16 model, single GPU)

Component	Before (training mode)	After (inference-only)
Model params (BF16)	54 GB	54 GB
DDP gradient buffers	54 GB	0
Adam m + v (FP32)	108 GB	0
Total	~216 GB OOM	~54 GB fits

Safety Verification

• Checkpoint key matching: load_checkpoint -> unwrap_model() handles DDP/non-DDP correctly via sharded_state_dict path
• eval + no_grad: forward_only() has @torch.no_grad() decorator; model_module.eval() is set
• Gradient isolation: compute_prm_teacher_log_probs output is .cpu() under @torch.no_grad(), never participates in training loss.backward()

Testing

• Model: Qwen3.5-27B (BF16)
• Config: PRM Teacher on 1x GPU (TP=1)
• Result: Model loads successfully, log-prob computation works correctly, no OOM