Fix nnz for scalars.

[hameerabbasi]

Addresses #47.

[mrocklin]

Is there a test that should have failed previously to this change? Can we add such a test to prevent regressions in the future?

[hameerabbasi]

I am hesitant to add a test (although I can add one) because indexing scalars in the future may give us scalars rather than COO objects. That's the way numpy/scipy.sparse do it, and that's how I believe it should be done. However, x[1, 1, ...] should give us a COO object for 2-D x.

[mrocklin]

OK. +1 from me then