Implement Transformer-M from GPS++

[DomInvivo]

https://arxiv.org/abs/2210.01765

Just the 3D input masking stuff

[DomInvivo]

To make it work, implement the part highlighted in Yellow in the image

Change the ATTENTION_LAYERS_DICT in file goli.nn.pyg_layers.gps_pyg to add the option transformer-M.

IMPORTANT: Make sure that you use the MultiheadAttentionMup, not the MultiheadAttention in the implementation. This will allow muTransfer to work as expected.