Edouard B.

I guess it makes sense to do a PR for both conv1d and selective_scan in the forward; there seems to be a faster than sequential implementation of the selective scan...