Add means for learned semantic position embeddings

[gkielian]

We should try to replicate the results of the following paper, but with a more learned approach like fire, and try to extend the results to other semantic spaces such as time series and NLP structures:

https://arxiv.org/pdf/2405.20671