Do not require default dtype if no dtypes in kind supported

[adityagoel4512]

The description of __array_namespace_info__'s default_dtypes method requires that you return a default dtype for all the keys specified.

The test suite here allows you to skip certain dtypes using the ARRAY_API_TESTS_SKIP_DTYPES environment variable. This means that it's possible to omit all the dtypes for a given kind. We use this in ndonnx to skip all complex dtypes since we're not able to support them at this time. In such a context, a default for "complex floating" doesn't make much sense.

I would like to only check the default_dtypes keys are present for kinds where we haven't skipped all the relevant dtypes. An alternative could be to make the value of the dictionary an optional dtype | None in a future version of the standard. I'd be happy with either option.

[ev-br]

This seems to overlap with #313, could you relate the approaches @adityagoel4512 @cbourjau ?

[cbourjau]

The PRs seem tangential, but I think both PRs make sense in their own right. In the light of skipping complex types with ARRAY_API_TESTS_SKIP_DTYPES it makes sense to test the other properties of __array_namespace_info__. However, an implementation without complex data types is not standard compliant, but that is not true for all data types that can be skipped with ARRAY_API_TESTS_SKIP_DTYPES. Could you add a new test case that checks for complex data types and fails regardless of ARRAY_API_TESTS_SKIP_DTYPES?

[adityagoel4512]

but that is not true for all data types that can be skipped with ARRAY_API_TESTS_SKIP_DTYPES.

What data type is this statement true for?

[cbourjau]

What data type is this statement true for?

I was under the impression that some of the "nonessential" data types were optional, but that is actually not the case. In that light, I think this PR looks good as it standards!

[ev-br]

Thanks @adityagoel4512