See: https://github.com/alugowski/fast_matrix_market

We already have mmread and mmwrite that use scipy.io, but fast_matrix_market ought to be faster (and better?). Let's make fast_matrix_market an optional dependency (it has Python wheels and is on conda-forge) and figure out the best way to use it. It has docs/example of creating GraphBLAS in C++ (see here), but I think we should be able to get by with the Python bindings.

One option for our API is to add engine="auto" keyword to these functions and try to use the fastest installed library by default.

Thanks @alugowski for creating fast_matrix_market and for bringing it to my attention.