https://mer.dev/posts/rethinking-attention-polynomial-alternatives-to-softmax-in-transformers/