Add weight dtype validation to fix CI issue (#2770)

# Motivation
pytorch/pytorch#172018 adds weight dtype
validation to CUDA and MPS, then introduces CI issue that we have
skipped in pytorch/pytorch#173335.
This PR aims to add weight dtype to XPU so that it resolves
pytorch/pytorch#173335

Author: guangyey

src/ATen/native/xpu/sycl/LossNLLKernel.cpp

@@ -406,6 +406,15 @@ void nll_loss_forward_kernel(
 
   auto weight_ = weight.defined() ? weight.contiguous() : weight;
 
+  if (weight_.defined()) {
+    TORCH_CHECK(
+        input.scalar_type() == weight_.scalar_type(),
+        "expected scalar type ",
+        input.scalar_type(),
+        " but found ",
+        weight_.scalar_type());
+  }
+
   if (reduction == at::Reduction::None && n_dims == 2) {
     at::native::resize_output(output, {batch_size});
     total_weight.zero_();