Implement Rotary Positional Embeddings

[aaprasad]

Rotary Positional Embeddings are a state of the art positional embedding that encodes both absolute and relative positional encodings using rotations. They seem to help a lot across a bunch of tasks from NLP to Computer Vision. The folks at EPFL have implemented in their tracking library and it could be useful for us to do the same. We should implement it as a function of the dreem.models.Embedding class similar to how we did with fixed/learned embeddings.

EDIT (Mustafa):
Add unit test in test_models.py