Fix the attention code to allow GPT-2 weight loading

[ramon-astudillo]

Upgrade the easier to understand GPT-2 attention code to allow loading GPT-2 weights.

i.e. avoid separate loaders/code for pre-trained and non pre-trained model weights https://github.com/LxMLS/lxmls-toolkit/blob/master/lxmls/transformers/model.py#L123

[israfelsr]

I'm workin in the branch unified-attention.
I managed to split the weights of the QKV projections but there is still something missing.
When prompting the model doesn't work correctly.

[israfelsr]

Done, we can close this one after merging unified-attention to master.