Fix illegal memory access when num indices = 0 in TBE inference (#1613)

Summary:
Pull Request resolved: #1613

The TBE inference kernel
(`*_split_embedding_codegen_forward*_kernel_small_L`) may illegally access
memory if the number of indices is zero.  This update ensures that the
TBE inference operator returns a zero-tensor when the number of
indices is zero, preventing the aforementioned problem.

Reviewed By: jianyuh

Differential Revision: D43588092

fbshipit-source-id: 95eb6b96c7b9cac8518072c03c6d382edfc47436

Author: sryap

fbgemm_gpu/codegen/embedding_forward_quantized_split_template.cu

@@ -667,17 +667,27 @@ Tensor int_nbit_split_embedding{{ "_nobag" if nobag else "" }}_codegen_forward_{
     if (o_dtype == SparseType::INT8) {
         total_adjusted_D += T * kINT8QparamsBytes;
     }
-    output = at::empty({B, total_adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    if (indices.numel() == 0) {
+      output = at::zeros({B, total_adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    }
+    else {
+      output = at::empty({B, total_adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    }
     {% else %}
     int64_t adjusted_D = D;
     if (o_dtype == SparseType::INT8) {
         adjusted_D += T * kINT8QparamsBytes;
     }
-    output = at::empty({total_L, adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    if (total_L == 0) {
+      output = at::zeros({total_L, adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    }
+    else {
+      output = at::empty({total_L, adjusted_D}, dev_weights.options().dtype(getScalarType(o_dtype)));
+    }
 
     {% endif %}
 
-    if (B == 0) {
+    if (B == 0 || indices.numel() == 0) {
       return output;
     }