speed up resolution matrix-vector multiplies

[sbailey]

See desihub/fastspecfit#181 which "Reimplement[s] enough of desispec.resolution.Resolution to perform matrix-vector operations, which is all we do with it, with a significant speedup via Numba compared to the scipy.sparse implementation."

Redrock doesn't directly use desispec.resolution.Resolution in order to avoid a desispec dependency in the core code, but it is still based upon scipy.sparse, including caching of various underlying formats for optimization. This might be even faster. Consider whether any of this fastspecfit PR could/should be ported to Redrock.