Omit Normalization Layer from MLP Class

[ChristophReich1996]

This pull request fixes issue #1851. The unused normalization layer is omitted from the Mlp class. Usage of the norm_layer argument is also omitted. Both BEiT(-v2) and EVA(-02) models are not affected by this change. EVA(-02) uses the norm_layer argument but since all configurations are using the SwiGLU instead if Mlp existing model weights are not affected. BEiT(-v2) sets in all configurations scale_mlp=False, meaning norm_layer was set to None, thus, also no affect here on existing models/weights.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[rwightman]

@ChristophReich1996 my bad on communication, I was assuming this'd go the other way, call the .norm() in forward() in case it's set, as this is a modification as per NormFormer https://arxiv.org/abs/2110.09456 ... as you noted it's only used in combo with the swiglu right now, but can be used in a non-gated MLP too