Adjust MFU to account for FP8

[lessw2020]

From internal discussions, logging an issue around updating our MFU calculations so that if FP8 is used, we can generate an accurate MFU number.

Atm - FP8 replaces wq/wk/wv/wo in Attention, and w1/w2/w3 in the MLP.

Thus, need an adjusted calculation.

In addition, would like to correctly pull the proper MFU (fp8 or bf16) based on the training config being run so this is handled automatically for the user.

[raghukiran1224]

yes, this needs to be updated. The MFU computations for fp8 are too good to be true :)
CC: @lchu-ibm