add evaluation of PRM trained by TRL

[CJReinforce]

Add the evaluation of PRM trained by the PRMTrainer of TRL.

I reproduced Qwen2.5-Math-7B-PRM800K using the PRMTranier of TRL. The performance of the reproduced PRM evaluated by run_eval_prm_trl.py are:

gsm8k error acc: 46.9, correct acc: 96.4, f1: 63.1
math error acc: 55.9, correct acc: 82.0, f1: 66.5
olympiadbench error acc: 39.0, correct acc: 67.8, f1: 49.6
omnimath error acc: 34.4, correct acc: 66.8, f1: 45.4
ProcessBench average f1: 56.1

[CJReinforce]

In the 05769fa commit, I fixed a bug where tokens were being concatenated incorrectly. Now appending token of separator after token of each step, instead of concatenating separator string and each step string before tokenization. This modification aligns with the training process in TRL.

After fixing the bug, the performance of my reproduced PRM evaluated by run_eval_prm_trl.py are:

gsm8k error acc: 52.2, correct acc: 96.9, f1: 67.8
math error acc: 58.9, correct acc: 81.3, f1: 68.3
olympiadbench error acc: 45.1, correct acc: 64.3, f1: 53.0
omnimath error acc: 42.7, correct acc: 66.4, f1: 52.0
ProcessBench. Average F1: 60.3