Fix Pre-LN for GPT2 Model

[flxst]

This PR fixes the pre-LN by reverting the changes (accidentally?) introduced here

References:

• Fig.1 & Tab. 1 in the paper On Layer Normalization in the Transformer Architecture
• nanoGPT