Unify the cuda and hip path for fbgemm in trainer embedding pruning (#2137)

Summary:
Pull Request resolved: #2137

Unifying the hip and cuda path so that we don't have to diverge in the user land. In addition, absorbed D58226889 and fixed the compat issue.

Reviewed By: jimmyyou, Yuzhen11, joebos

Differential Revision: D58775111

fbshipit-source-id: 23873deb59cd63c93b19fe597c3b22a50016d11a

Author: xw285cornell

torchrec/modules/itep_modules.py

@@ -17,14 +17,9 @@
 from torchrec.sparse.jagged_tensor import _pin_and_move, _to_offsets, KeyedJaggedTensor
 
 try:
-    if torch.version.hip:
-        torch.ops.load_library(
-            "//deeplearning/fbgemm/fbgemm_gpu:intraining_embedding_pruning_hip"
-        )
-    else:
-        torch.ops.load_library(
-            "//deeplearning/fbgemm/fbgemm_gpu:intraining_embedding_pruning_cuda"
-        )
+    torch.ops.load_library(
+        "//deeplearning/fbgemm/fbgemm_gpu:intraining_embedding_pruning"
+    )
 except OSError:
     pass