[CUDA][Convolution] Add missing launch bounds to vol2col_kernel (#1…

…28740)

Fix "too many resources requested" that can happen with recent toolkits on V100.

Pull Request resolved: #128740
Approved by: https://github.com/mikaylagawarecki

aten/src/ATen/native/cuda/vol2col.cuh

@@ -14,6 +14,7 @@ using namespace at::cuda::detail;
 
 // Kernel for fast unfold+copy on volumes
 template <typename T>
+C10_LAUNCH_BOUNDS_1(1024)
 __global__ void vol2col_kernel(
     const int64_t n,
     const T* data_vol,