Fix localizer multi-GPU failure (model.to vs device_map)

[KartikP]

On multi-GPU scoring tiers, HuggingfaceSubject loads models with device_map='auto' (sharded across GPUs). The localizer's extract_representations then called model.to(device), which tries to consolidate the sharded model onto a single GPU:

• 8B (fp32 ~32GB): OOMs trying to move the whole model onto one ~22GB A10G.
• 4B: partial consolidation leaves layers split, then the forward pass raises RuntimeError: Expected all tensors to be on the same device (cuda:0 and cuda:1).

Both surfaced as scoring failures on the 4-GPU medium tier (jenkins run 190); 0.6B/1.7B were unaffected because they land on the single-GPU small tier (no device_map, so the .to() is a no-op).

Fix: skip model.to(device) when the model is already dispatched via device_map (hf_device_map set). Inputs are already sent to self.device (the input-embedding shard), so the sharded forward works and 8B can use all GPUs instead of OOMing on one. Single-GPU loads are unchanged.