TST: ndimage: avoid implicit dtype conversions in vectorized_filter tests

This tests does this:  `vectorized_filter(input, xp.sum, output)`, with output
being `xp.empty_like(input)`. This is problematic for short integers because
- the output of xp.sum(int8_array) is int64 (the default int dtype), and
- internally, vectorized_filter does roughly,
`output[some_indices] = xp.sum(input[some_indices])`

So what happens here is that the output of the `sum` in the r.h.s. gets silently
downcast to the dtype of the l.h.s. by `__setitem__`.

This kind of sort of works in many array libraries, but is technically unspecified
by the Array API standard. And indeed this starts failing with array-api-strict
in data-apis/array-api-strict#157

Thus, a test only change to make downcasting explicit.

Author: ev-br

scipy/ndimage/tests/test_filters.py

@@ -2864,7 +2864,14 @@ def test_dtype_batch_memory(self, dtype, batch_memory, use_footprint, xp):
             assert res.dtype == sum_dtype
 
             output = xp.empty_like(input)
-            res = ndimage.vectorized_filter(input, xp.sum, output=output, **kwargs)
+            res = ndimage.vectorized_filter(
+                input,
+                lambda x, *args, **kw: xp.astype(
+                    xp.sum(x, *args, **kw), x.dtype, copy=False
+                ),
+                output=output,
+                **kwargs
+            )
             xp_assert_close(res, xp.astype(xp.stack(ref), dtype))
             assert res.dtype == dtype