Improve PRM inference speed

[CJReinforce]

This pr serves a similar purpose as this, in order to increase the speed of prm evaluation. But instead of modifying the content of the conversation (which can lead to inconsistencies of data format between evaluation and training time), I tried to infer all the process rewards in a single forward pass by concatenating the token of each part of the conversation and recording the position of the token needed to predict reward. This implementation does not modify the content of conversion in any way, ensuring consistency with the previous eval approach.

I tested it and the time taken to evaluate Deepseek-PRM-Deepseek-MATH500 using 4 A100 GPUs was 3 hours and 36 minutes.