Sparse COO divide by zero is zero

Thanks so much for creating this great library!

There seems to be an inconsistency when dividing by zero with sparse matrices:

```python
import dask
import dask.array as da
import sparse
print(f"dask version {dask.__version__}")
print(f"sparse version {sparse.__version__}")

coo = da.zeros(10).map_blocks(sparse.COO)

print((1 / coo).compute().todense())
# [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
# should be: [inf inf inf inf inf inf inf inf inf inf] ?

print((1 / coo.compute()).todense())
# [inf inf inf inf inf inf inf inf inf inf]
```
output:
```
dask version 2.1.0
sparse version 0.7.0
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[inf inf inf inf inf inf inf inf inf inf]
```

Am I missing something here?

Start by running the reproducer in the issue with the shown Dask and sparse versions, then compare the lazy map_blocks path with direct sparse.COO computation. Done means division by zero through the Dask path matches the direct path and returns the expected infinity values.

Written by the indexing model from the issue text.