MAINT: should not be using np.float64() on arrays

[andyfaff]

This is related to #18781.
When performing integrate.simpson(np.array([1,2,3.]), x=np.array([1, 2, 3.])) in a script I was experiencing DeprecationWarnings of the form reported #18781. The warning is also emitted with length 4 arrays.
These were not visible in the scipy test suite, probably because this line meant that all DeprecationWarnings, including the numpy warnings, weren't visible.

The numpy warnings were caused by the lines:

h0 = np.float64(h[sl0])
h1 = np.float64(h[sl1])

If h[sl0]) has length 1, then the np.float64 call raises the warning (similar argument for h1). The warning doesn't occur if the length is greater than 1. The solution is to use np.ndarray.astype. h[sl0] will always be an array, so casting to float should always work. The test suite already examines a case where the integrate set has length 4.

[mdhaber]

Thanks, @andyfaff.