These look pretty new, I see 240 failures locally on x86_64 Linux via python dev.py test -j 32 -b all -t scipy/ndimage/tests. I'm not sure this is surprising based on the discussion in gh-20809, but I'm using CuPy cupy-cuda12x 13.2.0. Bumping up to 13.3.0 (latest stable) has no effect either.

This causes the problem to go away, presumably by skipping CuPy, though perhaps too aggressively:

--- a/scipy/ndimage/tests/test_filters.py
+++ b/scipy/ndimage/tests/test_filters.py
@@ -25,7 +25,7 @@ skip_xp_backends = pytest.mark.skip_xp_backends
 xfail_xp_backends = pytest.mark.xfail_xp_backends
 pytestmark = [array_api_compatible, pytest.mark.usefixtures("skip_xp_backends"),
               pytest.mark.usefixtures("xfail_xp_backends"),
-              skip_xp_backends(cpu_only=True, exceptions=['cupy', 'jax.numpy']),]
+              skip_xp_backends(cpu_only=True, exceptions=['jax.numpy']),]

To state the obvious, that block is pretty hard to read. Anyway, could we skip these tests or something for now, that's a lot of noise in the array API suite.