BUG: sparse.csgraph, array types: support non-zero fill_values

[mtsokol]

Connected to: #20763, pydata/sparse#685, finch-tensor/finch-tensor-python#59

Hi @hameerabbasi,

This PR allows to pass pydata/sparse arrays with np.inf fill-values to the selected scipy.sparse.csgraph functions.

[mtsokol]

I think once pydata/sparse maintenance release is made then CI should be green.

[hameerabbasi]

The tests at scipy/sparse/csgraph/tests/test_pydata_sparse.py might need adjusting -- maintainence release is up on PyPI: https://pypi.org/project/sparse/0.15.4/

[hameerabbasi]

Thanks for the work on this @mtsokol!

[rgommers]

The test and changes LGTM, except for relying on a feature that is brand new in pydata sparse. How about wrapping the to_scipy_sparse(accept_fv) call in a try-except, catching the TypeError and then callign again without use of accept_fv? That keeps the default behavior with 0 fill value supported for older versions.

[hameerabbasi]

A few more places where a new sparse version is required.

[hameerabbasi]

Thanks for all the hard work here @mtsokol!

[rgommers]

Some comments above each try-except would be useful. I suggest:

# The `accept_fv` keyword is new in PyData Sparse 0.15.4 (May 2024),
# remove the `except` once the minimum supported version is >=0.15.4

[mtsokol]

Sure! I added comments.

[rgommers]

LGTM now, let's give this a go. Thanks @mtsokol, and thanks @hameerabbasi for the reviews.

[WarrenWeckesser]

It looks like the new code doesn't handle sparse versions before 0.15.4 correctly: #20861