Allow ndarray in elemwise again

[hameerabbasi]

I was considering allowing ndarray under elemwise again, provided that f(ndarrays, zeros_in_place_of_COO) is all-zero.

Upsides:

• No conversion required
• No matching/sorting required for the ndarray inputs at all. This makes it significantly faster for mixed sparse-dense elemwise.

Downsides:

• We'll have to broadcast all ndarrays against each other to actually test for densification. This could be huge. Or we can skip this step and risk wrong results.
  • Not so bad considering in almost all use-cases, the ndarrays will be tiny compared to COO.
• Puts a slight (but not huge) roadblock for arbitrary fill-values.

[mrocklin]

Meta-comment: My instinctual reaction here is "I don't know" and "we should really get more use cases so that we have a better metric by which to make these decisions".

[hameerabbasi]

In that case, I'll leave this open until such time as we can make these decisions.

[mrocklin]

Working on this seems fine to me, I'm mostly just thinking about things we can be doing that are more social and less technical. Maybe generalize CSR and target scipy devs?

[hameerabbasi]

Yes, I'm working on that PEP today. I hope to be done with the PEP soon.

I want things to be in a state that SciPy devs would agree to it. I could push my work to the repo, and if you're interested, you can contribute too. I've sent you an invite.

[hameerabbasi]

I'm blocked on this by numba/numba#2560 or #126.