Loss curve correctness for Llama 3.1 405B

[bhavya01]

🚀 Test loss curves against a reference implementation like huggingface for a large number of steps like 1000 steps.

The reference loss should come from a separate implementation (It can come from running huggingface models on GPUs or TPUs). Probably generate a loss file by running a cron job weekly and compare our implementation against the loss curve nightly for any regressions.