Add Exceptions to LanguageModel and question on DenseBlock Impl

[devindkim]

Adding some exception handling to LanguageModel and added a comment around the implementation of DenseBlock. Usually the widening -> gelu -> projection is sequential but this implementation isn't, I'm curious whether this is an intentional detail?

[snapsl]

I am kinda curious.. Does this degrade model performance?

[devindkim]

how the turn tables