Use gpu_library_selector for permute_pooled_embedding_ops_gpu (#1799)

Summary:
X-link: pytorch/FBGEMM#2435

Pull Request resolved: #1799

Using gpu_library_selector can avoid divergence in the usercode

A few notable changes:
* the original permute_pooled_embedding_ops_gpu is missing link_whole
* Make permute_pooled_embedding_function.cpp as its own library. This is because permute_pooled_embedding_function.cpp is included in both permute_pooled_embedding_ops_gpu and permute_pooled_embedding_ops_cpu. So when we link the _cpu and _gpu together, it'll complain about duplicate symbol.

Reviewed By: houseroad

Differential Revision: D54986292

fbshipit-source-id: b029c689eaf5f143b6d0b710aa30a06d1dd141f5

Author: xw285cornell

torchrec/distributed/utils.py

@@ -35,20 +35,9 @@
     torch.ops.load_library(
         "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_cpu"
     )
-    try:
-        if torch.version.hip:
-            torch.ops.load_library(
-                "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_gpu_hip"
-            )
-        else:
-            torch.ops.load_library(
-                "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_gpu_cuda"
-            )
-    except OSError:
-        # For backward compatibility
-        torch.ops.load_library(
-            "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_gpu"
-        )
+    torch.ops.load_library(
+        "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_gpu"
+    )
 except OSError:
     pass