A way to reinterpret a slice of size N as a multidimensional array with strides that are a factorization of N, including optional reverse order strides. Basically, do the stride bookkeeping internally so I can write an algorithm only considering the logic and optimize the striding order independently.
That's where you end up after heavy bikeshedding. Lots of features, terrible performance, as the OP points out.
I agree with you on sparse arrays and multidimensional slices, but this is basically the same as what you'd do manually. Saying that "track strides for me" is "lots of features" is a bit uncharitable.