fix: guard against division by zero in GPTRewardModel with empty batches

[themavik]

Summary

GPTRewardModel.forward() crashes with a ZeroDivisionError when the input batch has 0 or 1 samples:

bs = input_ids.shape[0] // 2  # 0 when shape[0] < 2
...
loss = loss / bs  # ZeroDivisionError

Additionally, torch.stack(chosen_end_scores) fails with RuntimeError on the empty list when bs == 0.

Changes

Add an early return guard after the batch size calculation. When bs == 0, the function returns:

• loss: zero tensor on the input device
• chosen_end_scores: empty tensor on the input device
• rejected_end_scores: empty tensor on the input device

Test Plan

• Batch with 2+ samples (normal paired input): behavior unchanged
• Batch with 1 sample: returns zero loss and empty scores instead of crashing
• Batch with 0 samples: returns zero loss and empty scores instead of crashing

Fixes #609