Add rotary embedding to GPT2 encoder

[Francesco215]

the nn.Embedding() function used to do learned positional encoding only works if the input sequence lenght is finite.

Adding rotary positional embedding will provide a way to give positional embedding to arbitrarly far away tokens

[Francesco215]

This repository implements rotary embeddings but uses a different convention for the indicies of Q, K and V

[Francesco215]

done #6