Change the TBE bounds check to match the TBE implementation. (#3773)

Summary:
X-link: facebookresearch/FBGEMM#853


TBE implementation accepts weights to be zero sized to represent empty weights.

Reviewed By: sryap

Differential Revision: D70660808

Author: Yavuz Yetim

fbgemm_gpu/codegen/utils/embedding_bounds_check_host_cpu.cpp

@@ -96,7 +96,7 @@ void bounds_check_indices_cpu(
               " is not equal to B (" + std::to_string(B) + ") * T (" +
               std::to_string(T) + ") + 1");
     }
-    if (weights.has_value()) {
+    if (weights.has_value() && weights->numel() != 0) {
       TORCH_CHECK(
           weights.value().size(0) == num_indices,
           "weights size " + std::to_string(weights.value().size(0)) +

fbgemm_gpu/codegen/utils/embedding_bounds_check_v1.cu

@@ -227,7 +227,7 @@ void _bounds_check_indices_cuda_v1(
             " is not equal to B (" + std::to_string(B) + ") * T (" +
             std::to_string(T) + ") + 1");
   }
-  if (weights.has_value()) {
+  if (weights.has_value() && weights->numel() != 0) {
     TORCH_CHECK(
         weights.value().size(0) == num_indices,
         "weights size " + std::to_string(weights.value().size(0)) +