Convert self object manually to np.ndarray on fallback

[manopapad]

Otherwise we get errors like the following when we try to invoke an unimplemented array-level method:

TypeError: descriptor 'std' for 'numpy.ndarray' objects doesn't apply to a 'ndarray' object

[bryevdv]

Would isinstance check with types.MethodType work instead of having to pass an explicit flag?

[bryevdv]

@manopapad this change also seems to work, without needing to change API:

diff --git a/cunumeric/coverage.py b/cunumeric/coverage.py
index 2056bbed..07b3f72c 100644
--- a/cunumeric/coverage.py
+++ b/cunumeric/coverage.py
@@ -150,6 +150,8 @@ def unimplemented(
                 location=location,
                 implemented=False,
             )
+            if isinstance(func, (MethodType, MethodDescriptorType)):
+                args = (args[0].__array__(),) + args[1:]
             return func(*args, **kwargs)
 
     else:
@@ -162,6 +164,8 @@ def unimplemented(
                 stacklevel=stacklevel,
                 category=RuntimeWarning,
             )
+            if isinstance(func, (MethodType, MethodDescriptorType)):
+                args = (args[0].__array__(),) + args[1:]
             return func(*args, **kwargs)
 
     wrapper._cunumeric = CuWrapperMetadata(implemented=False)

Let me open a PR with this and also try to add a test for it.

[bryevdv]

Although, either of the changes means that clone_class is now only useful for ndarray, since it now implicitly assumes all self args should be arrays. If we want clone_class to retain any semblance of being a generic tool, we will have to find a way not to hard-code this codepath unconditionally for any and every method.

[bryevdv]

Ugh. Testing this is a mess. We can't just make some toy class in the test, because the numpy gets unhappy. But if we base tests on real ndarray then things that are unimplemented today, may be implemented in the future, negating the the test.

[manopapad]

Superseded by #431