wrapper subclasses: support non-cpu device for dynamic shape overload (#107926)
This is tested by AOTAutograd later in the stack, but I can add direct tests if anyone wants them.
Previously, the second overload of `_make_wrapper_subclass` (which supports dynamic shapes) would just always return a wrapper tensor that reported as being on `cpu`. This updates it to properly respect the `device` arg that was passed in.
At first I thought about doing this the same way that FakeTensor does it (override device to do a custom impl), but that seemed overly complicated. Since the subclass is a wrapper, we can just bake in the value on the wrapper.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107926
Approved by: https://github.com/ezyang
ghstack dependencies: #107915, #107916
