Fix int32 overflow in max_pool3d backward kernel

Widen stride variables in `MaxPool3dBackwardKernelFunctor` from `int` to
`int64_t` to prevent overflow when `batch * in_nbatch_stride` exceeds
`INT_MAX`. This caused incorrect gradient writes for large tensors
(e.g. shape `[70, 32, 100, 100, 100]`).

Fixes test case: `test_pool3d_large_size_int64`

Author: pbielak

src/ATen/native/xpu/sycl/DilatedMaxPool3d.cpp

@@ -347,10 +347,10 @@ struct MaxPool3dBackwardKernelFunctor {
       const int64_t* indicesData,
       int features,
       int64_t gradOutputSize,
-      int out_cf_channel_stride,
-      int in_cf_channel_stride,
-      int out_nbatch_stride,
-      int in_nbatch_stride)
+      int64_t out_cf_channel_stride,
+      int64_t in_cf_channel_stride,
+      int64_t out_nbatch_stride,
+      int64_t in_nbatch_stride)
       : gradInputData_(gradInputData),
         gradOutputData_(gradOutputData),
         indicesData_(indicesData),
@@ -367,10 +367,10 @@ struct MaxPool3dBackwardKernelFunctor {
   const int64_t* indicesData_;
   int features_;
   int64_t gradOutputSize_;
-  int out_cf_channel_stride_;
-  int in_cf_channel_stride_;
-  int out_nbatch_stride_;
-  int in_nbatch_stride_;
+  int64_t out_cf_channel_stride_;
+  int64_t in_cf_channel_stride_;
+  int64_t out_nbatch_stride_;
+  int64_t in_nbatch_stride_;
 };
 
 template <typename scalar_t, bool channels_last>
@@ -386,12 +386,13 @@ void max_pool3d_with_indices_backward_template(
     int otime,
     int oheight,
     int owidth) {
-  int64_t gradOutputSize = obatch * features * otime * oheight * owidth;
+  int64_t gradOutputSize =
+      (int64_t)obatch * features * otime * oheight * owidth;
 
-  auto out_cf_channel_stride = otime * oheight * owidth;
-  auto in_cf_channel_stride = itime * iheight * iwidth;
-  auto out_nbatch_stride = features * out_cf_channel_stride;
-  auto in_nbatch_stride = features * in_cf_channel_stride;
+  int64_t out_cf_channel_stride = (int64_t)otime * oheight * owidth;
+  int64_t in_cf_channel_stride = (int64_t)itime * iheight * iwidth;
+  int64_t out_nbatch_stride = (int64_t)features * out_cf_channel_stride;
+  int64_t in_nbatch_stride = (int64_t)features * in_cf_channel_stride;
   MaxPool3dBackwardKernelFunctor<scalar_t, channels_last> kfn(
       gradInputData,
       gradOutputData,