[WIP] Sparse pairwise semi-ring kernel

[cjnolet]

This PR implements Lp-style distances and their derivatives with a matrix-vector-operation-style mapping to warps and a semi-ring-style pluggability w/ accumulation and reduction functions. The initial version (in this PR) will only aim to provide an Lp distance primitive that shows a significant speedup over CPU.

Assuming it will be fast enough, future versions of this implementation could potentially have several benefits even over the existing dot product (and expanded euclidean) prims, which use cusparse csrgemm:

1. The csrgemm in cusparse that accepts both sparse inputs outputs a sparse matrix, so we need to allocate double the memory and waste time converting it to dense.
2. The B input to the csrgemm needs to be transposed. Unfortunately, CSC is not able to share any data w/ a CSR, so that's even more extra allocations.
3. This pairwise distance kernel will enable more than just a matrix multiply, it can actually be used to implement all the Lp and inner-product-based variants. The cusparse GEMM is good at the one thing, and we're still left to our own implementations for any Lp variants aside from expanded euclidean.
4. Today, we are still stuck supporting the 32-bit indexing API from cusparse, and even with the 64-bit indexing API, there are still many functions that only support 32-bit. This will allow us to support 64-bit natively, just through type templates.