These warning filters are not threadsafe, while it is nicer if the
comparison functions can be called in a threaded environment since
they do no explicite warning checking.

Less filters are also cleaner in general, though may also mean that
downstream projects see warnings that were previously hidden.

gh-8410 breaks a large number of astropy tests, because it sets up
a boolean array for values that should actually be compared (i.e.,
are not `nan` or `inf`) using `zeros_like`. The latter means that
for subclasses, the boolean test array is not a plain `ndarray` but
the subclass. But for astropy's `Quantity`, the `all` method is
undefined.

This commit ensures the test arrays from `isinf` and `isnan` are
used directly.

This failed to raise AssertionError in an earlier rewrite of
assert_array_compare and is thus tested explicitly now.

Added tests for simple arrays/elements,
tests that ensure that nans are not compared when they are in the same
position in both arrays but raise AssertionError when they are not,
and tests that ensure inf logic
(-np.inf < np.inf, 1.0 < np.inf, -np.inf < 1.0)
and that all other cases fail.

This means that warnings of different origin then the one tested for
behave normally. If the normal behaviour is to igonre them this might
decrease specificity in rare cases. In most cases the specificity
will be slightly higher.

Also modify the corresponding test to suppress the non Deprecation
warnings created to test specificity.