Update usages of torch.library APIs

We deprecated impl_abstract. This PR replaces it with the new API
(register_fake). register_fake also (sometimes) requires a
`set_python_module` in C++, so I add that as well.

Test Plan:
- existing tests

torchvision/_meta_registrations.py

@@ -160,7 +160,7 @@ def meta_ps_roi_pool_backward(
     return grad.new_empty((batch_size, channels, height, width))
 
 
-@torch._custom_ops.impl_abstract("torchvision::nms")
+@torch.library.register_fake("torchvision::nms")
 def meta_nms(dets, scores, iou_threshold):
     torch._check(dets.dim() == 2, lambda: f"boxes should be a 2d tensor, got {dets.dim()}D")
     torch._check(dets.size(1) == 4, lambda: f"boxes should have 4 elements in dimension 1, got {dets.size(1)}")

torchvision/csrc/ops/nms.cpp

@@ -19,6 +19,7 @@ at::Tensor nms(
 }
 
 TORCH_LIBRARY_FRAGMENT(torchvision, m) {
+  m.set_python_module("torchvision._meta_registrations");
   m.def(TORCH_SELECTIVE_SCHEMA(
       "torchvision::nms(Tensor dets, Tensor scores, float iou_threshold) -> Tensor"));
 }