Add "full" gelu without approximation

[se-schmitt]

• adds the activation function gelu_noapprox which calculates the full gelu (pytorch default case, see here).

Context: I have implemented a thermodynamic model based on a ChemBERTa model here. To obtain the same results (within numerical accuracy) as the original model (which was implemented using pytorch, see here), the "full" gelu activation function is needed (locally tested). So I would be happy to have this feature added in Transformers.jl, so that I can submit the corresponding changes to Clapeyron.jl.

[se-schmitt]

@chengchingwen have you had a chance to look at this?

[chengchingwen]

If the gelu without approx is the default in pytorch, we should also think about having that implemented in NNlib.jl. The question here is that there is no "gelu_noapprox" in huggingface/transformers so it wouldn't be loaded automatically unless you manually update the config.

So I would suggest:

1. open an issue/PR in NNlib.jl with the gelu implemented with erf
2. instead of adding "gelu_noapprox" to ACT2FN, directly replace the gelu for [gelu, gelu_python].
3. It would be better to have a small test testing the gelu values and gradients