ndimage.convolve should take weights' and cval's dtypes into account when determining the output dtype

[anntzer]

If ndimage.convolve is passed input as integer array and weights or cval (with mode="constant") as floating point, the output buffer should be floating point, not integer.

Reproducing code example:

below the examples are with integer weights and floating point cval, but the issue is also present with just floating point weights.

In [8]: scipy.ndimage.convolve([1, 2, 3, 2, 1], [1, 1, 1], mode="constant", cval=0.1)                               
Out[8]: array([3, 6, 7, 6, 3])

In [9]: scipy.ndimage.convolve([1., 2., 3., 2., 1.], [1, 1, 1], mode="constant", cval=0.1)                          
Out[9]: array([3.1, 6. , 7. , 6. , 3.1])

and more spectacularly:

In [10]: scipy.ndimage.convolve([1, 2, 3, 2, 1], [1, 1, 1], mode="constant", cval=np.nan)                           
Out[10]: 
array([-9223372036854775808,                    6,                    7,
                          6, -9223372036854775808])

In [11]: scipy.ndimage.convolve([1., 2., 3., 2., 1.], [1, 1, 1], mode="constant", cval=np.nan)                      
Out[11]: array([nan,  6.,  7.,  6., nan])

Scipy/Numpy/Python version information:

1.3.0 1.17.0rc1 sys.version_info(major=3, minor=7, micro=3, releaselevel='final', serial=0)