Adjust kNumThreads for bounds_check_indices_kernel (#3299)

Summary:
X-link: facebookresearch/FBGEMM#398

Pull Request resolved: #3299

Increase the thread block size to increase the number of warps per SM.
This results in a better kernel time

Reviewed By: spcyppt

Differential Revision: D65071345

fbshipit-source-id: 0196eadf38278373fe36edb470221c12f219510b

fbgemm_gpu/codegen/utils/embedding_bounds_check_v2.cu

@@ -251,7 +251,7 @@ void _bounds_check_indices_cuda_v2(
             " is not equal to indices size " + std::to_string(num_indices));
   }
 
-  constexpr size_t kNumThreads = 256;
+  constexpr size_t kNumThreads = 1024;
   const auto max_B_ = vbe ? max_B : B;
 
   const int32_t vbe_bound = max_B_ * T;