BUG: sparse: Propagate dtype through CSR/CSC constructors

[perimosocordiae]

Reference issue

Fixes gh-13329

What does this implement/fix?

Propagates the dtype parameter through the intermediate COO format constructor.

[tylerjereddy]

@perimosocordiae When I run this feature branch with the original example code in the matching issue I still get a traceback (below). Is there still something off in the original reproducer from @matteodellamico?

scipy.__version__: 1.7.0.dev0+7185d00
Traceback (most recent call last):
  File "test.py", line 21, in <module>
    assert v == data[i], (i, data[i], v)
AssertionError: (4, 14115858573662012539, 14115858573662013440)

[perimosocordiae]

Did you use the fixed reproducer? The original version could have duplicate entries.

[tylerjereddy]

I thought I did--I believe the author edited the code and I copy-pasted from there.

[WarrenWeckesser]

The test passes in the master branch, because 2**63 happens to be a number that survives the round trip through floating point. The big value used in the test should be changed to, say, 2**63 + 1.

The test will then fail, because the fix is not complete. The dtype has to be passed to the array function that is eventually called at

scipy/scipy/sparse/coo.py

Line 159 in 01838d2

self.data = np.array(obj, copy=copy)

[perimosocordiae]

Thanks for the double-checking @tylerjereddy and the explanation @WarrenWeckesser. I learned something new about the way numpy does dtype inference today:

In [1]: np.array([2**64 - 1]).dtype                                                                                                      
Out[1]: dtype('uint64')

In [2]: np.array([2**64 - 1] * 5).dtype                                                                                                  
Out[2]: dtype('uint64')

In [3]: np.array([2**64 - 1, 2**32 - 1]).dtype                                                                                           
Out[3]: dtype('float64')

With my latest commit, this bug should actually now be fixed.

[WarrenWeckesser]

One small PEP-8 change, otherwise looks good.

[WarrenWeckesser]

Merged, thanks @perimosocordiae!