[rocm] Enable faster load of fused-sgd

[crcrpar]

Currently it's turned off to avoid the hang of test_cuda

          I don't have an explanation yet, but since the kernel had two potential paths (`use_faster_load_store`), I used a preprocessor #ifdef for rocm to force the other path.  This worked.  So there was something wrong with using the faster load store path.  This is my diff that worked for ROCm

diff --git a/aten/src/ATen/native/cuda/FusedSgdKernel.cu b/aten/src/ATen/native/cuda/FusedSgdKernel.cu
index a00b47ff79..0c52ee5783 100644
--- a/aten/src/ATen/native/cuda/FusedSgdKernel.cu
+++ b/aten/src/ATen/native/cuda/FusedSgdKernel.cu
@@ -87,6 +87,7 @@ struct FusedSgdMathFunctor {
         init_args<depth>(args, tl, chunk_idx, chunk_size, tensor_loc)};
     n -= chunk_idx * chunk_size;

+#ifndef USE_ROCM
     const auto use_faster_load_store =
         (n % kILP == 0) && (chunk_size % kILP == 0) && all_aligned;
     if (use_faster_load_store) {
@@ -116,7 +117,9 @@ struct FusedSgdMathFunctor {
           load_store(args[2], r_args[2], i_start, 0);
         }
       }
-    } else {
+    } else
+#endif
+    {
       for (auto i_start = 0; i_start < n && i_start < chunk_size;
            i_start += blockDim.x * kILP) {
         load_args<depth>(r_args, args, i_start, chunk_size, n);

Originally posted by @jeffdaily in #116585 (comment)

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang

[petrex]

@crcrpar Any particular unit tests we can use to validate the feature?