Add `intermediate_size` to GPT-NeoX models

[dtamayo-nlp]

The current implementation only allows to set intermediate_size for Llama models, but I would like to be capable to change the intermediate_size in GPT-NeoX models.

I have tested this implementation under a quick training, inference and conversion and it seems that it doesn't give any bugs. I hope it helps!

[CLAassistant]

All committers have signed the CLA.

[Quentin-Anthony]

This looks good. Need to make it consistent with mamba and RWKV. We also need some TODO statements about revisiting this once we add swiglu.

@jahatef is on it

[jahatef]

Added support for mamba and RWKV, and added TODOs.

[StellaAthena]

With these changes, is there a point to having separate Linear and LLaMA Linear definitions? At a glance it looks like all the differences are configurable, with the only difference being what is assumed as the default behavior if unspecified.

[jahatef]

Refactored the activations and MLP layer to get rid of our redundant llama mlp class, and added some activations functions from https://arxiv.org/pdf/2002.05202.