BUG: CubicSpline segfaults when passing empty values for `y`with numpy >=1.23

Describe your issue.

While the docs are clear and the values for y must be finite, one could pass empty array and, some iterative interpolations we had in our code that relied on that behavior, “worked.”

We fixed our code to return a ValueError now instead of relying on that behaviour. However, b/c the old code segfaults, maybe some checks should be in place in SciPy to avoid empty y values. Please feel free to close this issue if this doesn’t make sense.

Reproducing Code Example

from scipy.interpolate import CubicSpline
import numpy as np

# Segfaults on my machine when it reaches size>=10 but passes with <10.
x = np.arange(10)

y = np.empty(len(x)) * np.NaN
y_no_nans = y[~np.isnan(y)].reshape(y.shape[0], -1)

# In our code y_no_nans was slowly iterated to reach valid finite value and worked for numpy <1.23.
spline = CubicSpline(x, y_no_nans, bc_type='periodic')

Error message

Sometimes it segfaults, sometimes it "works" :-/

SciPy/NumPy/Python version information

1.9.1 1.23.3 sys.version_info(major=3, minor=10, micro=6, releaselevel=‘final’, serial=0)

[ev-br]

@ocefpaf heads up: https://github.com/scipy/scipy/pull/17305 is in, thank you for working on this!

[ev-br]

OK, https://github.com/scipy/scipy/pull/17305 implements the suggested API with an intent to close this issue. Reviews would be appreciated.