fix: add LoRA support and batch generation to gsm8k eval script

[yuxizhe]

This commit addresses two issues in the evaluation script:

1. Fix LoRA Application: I found out the evaluation script wasn't applying the LoRARequest properly. This meant we were just evaluating the base model every time, which is why the results looked almost the same as pre-training. Inquiry about the detailed config of reproduction #14
2. Implement Batch Evaluation: Added batched inference to the evaluation process. This optimization drastically improves performance, reducing the total evaluation time from several hours down to a few minutes.