nnz not correct for scalars.

The `x.nnz` should be `1` for for a nonzero scalar `x`, yet it is zero. MWE:

```
import numpy as np
x = 1 + np.random.rand(5, 5)
import sparse
xs = sparse.COO.from_numpy(x)
xs[1, 1]
Out[6]: <COO: shape=(), dtype=float64, nnz=0, sorted=True, duplicates=False>
```