Dependency refinement sparse, numba

[zklaus]

Currently, xESMF simply depends on sparse, with no lower bound, and does not depend on numba explicitly. In conda-forge/esmvaltool-suite-feedstock#5 we ran into a problem where that leads to numba missing because older sparse builds do not depend on it.

I suggest that xESMF depends on numba itself since it is used in smm.py and thus a direct dependency. I also wonder if you want to place a lower bound on the sparse dependency?

[raphaeldussin]

@zklaus that sounds like a good solution, what lower bound for sparse are you thinking of?

[zklaus]

@raphaeldussin, I don't know about the lower bound on sparse. My thinking was that a lower bound on sparse might pull in numba automatically, but also, I know that sparse has undergone substantial development and I am not sure which version is the oldest that you have done your tests with.

If you add the numba dependency (as I think you should because you are using it directly), there is no technical need for a lower bound on sparse. But if you know or suspect that you need a certain maturity from sparse, you may still consider it.