BUG: Ensure non-array logspace base does not influence dtype of output.

[mhvk]

Fixes gh-24957

[rgommers]

Thanks @mhvk!

The CircleCI failure is related:

numpy.logspace
--------------

File "venv/lib/python3.11/site-packages/numpy/__init__.py", line ?, in logspace
Failed example:
    np.logspace(2.0, 3.0, num=4, base=[2.0, 3.0], axis=-1)
Exception raised:
    Traceback (most recent call last):
      File "/home/circleci/.pyenv/versions/3.11.4/lib/python3.11/doctest.py", line 1351, in __run
        exec(compile(example.source, filename, "single",
      File "<doctest logspace[5]>", line 1, in <module>
        np.logspace(2.0, 3.0, num=4, base=[2.0, 3.0], axis=-1)
      File "/home/circleci/repo/venv/lib/python3.11/site-packages/numpy/_core/function_base.py", line 312, in logspace
        return _nx.power(base, y)
               ^^^^^^^^^^^^^^^^^^
    ValueError: operands could not be broadcast together with shapes (2,) (4,)

[mhvk]

OK, hopefully fixed now...

[mhvk]

The freebsd failure is unrelated, so this should be all OK now.

[rgommers]

LGTM, thanks for the fix @mhvk!

[seberg]

Thanks, it's not actually quite right, because we are not dealing with the other arguments being scalars. But at least it fixes the bigger problem.

[mhvk]

@seberg - ah, yes, I thought about doing it after the linspace call, but then one has to deal with the fact that axis gets applied differently... Probably would have been better to use your array_converter or so, and use that to calculate the dtype first.

[seberg]

Yeah, array-converter should help a bit, but it simply is tedious. The reason is that base can be the only array in which case you need to broadcast the linspace result manually.