Add code for evaluating pass @ k to inference_and_check

[erictang000]

Fixes minor bug in perform_check and adds code for checking pass @ k metric for n > 1 samples.

For example if we run the following with a saved file DeepSeek-R1-Distill-Qwen-7B_aime_train_None_False_0_-1.json with n=128 examples per question

python inference_and_check.py --model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --task aime  --split train --max_tokens 32768   --inference --n 128  --temperatures 0.6 --tp 1 --check

We will get the following output now:

Temperature: [0.6]
Loaded 30 existing results.
Found 3840 responses requiring reject sampling...
Processing Reject Sampling: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3840/3840 [00:02<00:00, 1432.24it/s]
Final reject-sampling accuracy: 2052/3840
Actual accuracy: 0.534375
Final pass @ k:
k: 128, pass @ k: 90.0
k: 64, pass @ k: 84.999
k: 32, pass @ k: 82.379
k: 16, pass @ k: 80.524
k: 8, pass @ k: 78.576
k: 4, pass @ k: 74.26
k: 2, pass @ k: 65.496
k: 1, pass @ k: 53.438
Temperature 0.6 acc: 27/30 (0.9)