[ROCm] Fix for a bug in ROCm batchnorm implementation.

Fixing the bug also makes the unit-test `//tensorflow/python/keras/layers:normalization_test_gpu` pass, so removing the `no_rocm` tag from it as well.

tensorflow/python/keras/layers/BUILD

@@ -715,7 +715,6 @@ cuda_py_test(
     python_version = "PY3",
     shard_count = 4,
     tags = [
-        "no_rocm",
         "notsan",
     ],
     xla_tags = [

tensorflow/stream_executor/rocm/rocm_dnn.cc

@@ -3584,8 +3584,6 @@ bool MIOpenSupport::DoBatchNormalizationForwardImpl(
 
   auto status = miopenStatusInvalidValue;
   if (is_training) {
-    stream->ThenMemZero(batch_mean, batch_mean->size());
-    stream->ThenMemZero(batch_var, batch_var->size());
     status = wrap::miopenBatchNormalizationForwardTraining(
         miopen.handle(), mode, &one, &zero, x_descriptor.handle(), x.opaque(),
         x_descriptor.handle(), y->opaque(), scale_offset_descriptor.handle(),