Use more correct B matrix for S6 equinox model

[smorad]

This modifies the equinox S6 model to more closely match how the original paper discretizes the B matrix: https://arxiv.org/pdf/2312.00752. The results are very similar, but we should be consistent with the paper. Increasing the dimensionality of dt also seems to help quite a bit, and so we set dt to be recurrent_size by default. As a result, we split the S6 model into the more aptly named S6 and S6D (diagonal) modules.