Backport of #21925.

This would be the minimal fix for gh-21911 (and an older issue somewhere). However, it would also set a precedence that subarray to object casts will create copies.
That is likely fine, since those casts usually crashed for a while and in many places a copy is OK. It was also always broken from an ownership perspective.
However, for code that used this and assigned to the result, things would obviously change.

There are three options:

1. Less-minimal fix, because we want view behavior here (and fix it).
2. Keep it as it is (it was so unreliable that failures are hopefully minimal, and why would one cast to object to then assign?).
3. We could tag on a warning, I don't think the warning "flag" is used right now. I.e. print out a UserWarning that this is a copy, but may have been a view whenever the array may be written to. (This may not be quite reliable for buffer users like cython, but not sure.)