Ads parallel MLPerf accuracy eval.

[patemotter]

For MLPerf accuracy runs we need to compute the ROUGE scores and compare against expected outputs.

This change adds a parallelized the code, which provides a speedup of 20x (10m -> 0.5m) when evaluating the 25k MLPerf dataset for Llama2-70B. Since this doesn't change the underlying computation of the scores we get the same results as the serial version.

Original Serial Results
{'rouge1': 44.3842, 'rouge2': 21.8723, 'rougeL': 28.5027, 'rougeLsum': 41.947, 'gen_len': 28135648, 'gen_num': 24576, 'gen_tok_len': 7171079, 'tokens_per_sample': 291.8}

real    10m4.329s
user    10m8.792s
sys     0m8.158s


Parallel Results
{'rouge1': 44.3842, 'rouge2': 21.8723, 'rougeL': 28.5027, 'rougeLsum': 41.947, 'gen_len': 28135648, 'gen_num': 24576, 'gen_tok_len': 7171079, 'tokens_per_sample': 291.8}

real    0m29.519s
user    13m2.545s
sys     2m43.586s

[vipannalla]

Can add a way (env_car or option in llama_offline_run.sh) to use this new and faster eval file?