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

torch/csrc/autograd/python_variable.cpp

@@ -702,7 +702,7 @@ static PyObject* THPVariable_make_wrapper_subclass(
     Storage storage{
         Storage::use_byte_size_t{},
         0,
-        at::DataPtr{},
+        at::DataPtr{nullptr, r.device(7)},
         /*allocator=*/c10::GetAllocator(c10::kMeta),
         /*resizeable=*/true};