Integrate LucidRain's RotaryEmbeddings

[suchenzang]

See https://github.com/lucidrains/rotary-embedding-torch/blob/main/rotary_embedding_torch/rotary_embedding_torch.py

And from PaLM paper:

We use RoPE embeddings (Su et al., 2021) rather than absolute or relative position embeddings, since RoPE embeddings have been shown to have better performance on long sequence lengths.

[erip]

Since you already have xformers as a soft dependency, you should be able to pull it in directly from there if it's installed (similar to flash attention).

[suchenzang]

It would be nice to have more configurability via:
https://github.com/lucidrains/rotary-embedding-torch/blob/6868f6ff30898989e4aa5890973911b2edc5e8d8/rotary_embedding_torch/rotary_embedding_torch.py#L60-L72

vs

https://github.com/facebookresearch/xformers/blob/bc08bbc631348913a3c37b4e09832973ff93a398/xformers/components/positional_embedding/rotary.py#L49-L57