fix: use untrainable precomputed slope

fix: adjust example to fit default `d_model`

fix: make residual preln explicit in encoder layer

docs: clarify encoder decoder usage

fix: apply norm before kqv

docs: add illustration link

docs: populate readme

fix: add assert for `max_len` check

chore: rename `models.py` -> `model.py`

feature: implement alibi

Initial commit