API to change fill-value

Hi!

This is an awesome package and I'm having fun using it to its full potential. Hopefully this is the right place for this. I'm having an issue where, after a series of boolean operations, my fill value becomes True and my data value becomes False. This limits my ability to do further calculations.

I have a sparse, boolean n x n x m array.

>>> import sparse as sp
>>> m = sp.COO(
...    coords=np.array([[0, 0, 0], [0, 0, 1],
...                     [0, 1, 0],
...                     [1, 0, 0],
...                     [1, 1, 0], [1, 1, 1],
...                     [2, 2, 1], 
...                     [3, 3, 0], [3, 3, 1]
...                     ]).T,
...    data=1,
...    shape=(4, 4, 2),
... )
>>> m
<COO: shape=(4, 4, 2), dtype=int64, nnz=9, fill_value=0>

I want to identify rows/columns where there is not any value that is present in all three coordinates by checking for the number of positions that have something present and then that this is less than the number of regions.

>>> pres = (m.sum(axis=1) > 0).sum(axis=1) < 2
>>> pres
<COO: shape=(4,), dtype=bool, nnz=3, fill_value=True>
>>> pres.todense()
array([False, False,  True, False])

If I try to do something like flip my fill value, then I just get a matrix of zeros

>>> pres.fill_value = 0
>>> pres
<COO: shape=(4,), dtype=bool, nnz=3, fill_value=0>
>>> pres.todense()
array([False, False,  False, False])

But, the fact that my array density flips seems like a possible bug, or at least a parameter I'd like more control over.

Reproduce the reported COO example and trace how boolean reductions determine fill_value and stored data. Review the existing COO fill_value behavior and related tests or API entry points; done means a supported way to change the fill value preserves the correct dense result without turning valid values into zeros.