Add `__complex__` to array object

We currently have respective dunder methods for bool/ints/floats, so if just for consistency x.__complex__() seems nice. Personally I’ve found the other methods hella useful for array-api-tests and Hypothesis, so imagine this will prove useful too.

I think we’d want to specify this just for complex-valued floating-point arrays, and adopters can choose to support real-valud float arrays if-and-however they so wish.

[leofang]

I think the story for __complex__ is different. In other places, we strive to avoid complex-to-real conversions, but here with __complex__ the direction is different (whatever-to-complex), so I am less concerned.

[seberg]

IMO it must work for all 0-D inputs covered by the standard (numerical and boolean). Two reasons:

1. That is how Python does it.
2. This would be related to casting and thus: real_arr.astype(complex_dtype). That I think must work? (This is not promotion: there are no mixed dtypes involved.)

I could see a point against boolean, but it seems no worse than for int/float, so unless that is exclused there, it should not be here I think.