[MPS] Fix `test_embedding_dense_backward` (#88847)

By copying randomly initialized weights distribution from MPS `nn.Embedding` to `cpu`

Test plan: `python test_mps.py -k test_embedding_dense_backward --repeat 150`

Fixes https://github.com/pytorch/pytorch/issues/88679

Pull Request resolved: https://github.com/pytorch/pytorch/pull/88847
Approved by: https://github.com/seemethere

test/test_mps.py

diff --git a/test/test_mps.py b/test/test_mps.py
index 2ff5a9d..30546f5 100644
--- a/test/test_mps.py
+++ b/test/test_mps.py
@@ -4282,8 +4282,9 @@
     def test_embedding_dense_backward(self):
         def helper(n, d, m, idx):
             embeddingMPS = nn.Embedding(n, d, max_norm=True, device='mps')
+            emedding_weight = embeddingMPS.weight.detach().cpu()
             W_MPS = torch.randn((m, d), requires_grad=True, device='mps')
-            idx_MPS = torch.tensor(idx).to('mps')
+            idx_MPS = torch.tensor(idx, device='mps')
             a_MPS = embeddingMPS.weight.clone() @ W_MPS.t()  # weight must be cloned for this to be differentiable
             a_MPS.retain_grad()
             b_MPS = embeddingMPS(idx_MPS) @ W_MPS.t()  # modifies weight in-place
@@ -4292,7 +4293,7 @@
             loss_MPS = out_MPS.sigmoid().prod()
             loss_MPS.backward()
 
-            embeddingCPU = nn.Embedding(n, d, max_norm=True, scale_grad_by_freq=True)
+            embeddingCPU = nn.Embedding(n, d, max_norm=True, _weight=emedding_weight)
             W_CPU = W_MPS.to('cpu')
             idx_CPU = torch.tensor(idx)
             a_CPU = embeddingCPU.weight.clone() @ W_CPU.t()  # weight must be cloned for this to be differentiable