OP pt: add contiguous in grad output (deepmodeling#3917)

deepmodeling#3910

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

- **Bug Fixes**
- Improved tensor handling in the `Border` class to ensure proper memory
management and performance by making tensors contiguous.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

source/op/pt/comm.cc

@@ -197,7 +197,7 @@ class Border : public torch::autograd::Function<Border> {
     torch::Tensor nlocal_tensor = saved_variables[6];
     torch::Tensor nghost_tensor = saved_variables[7];
 
-    torch::Tensor d_local_g1_tensor = grad_output[0];
+    torch::Tensor d_local_g1_tensor = grad_output[0].contiguous();
 #ifdef USE_MPI
     int mpi_init = 0;
     MPI_Initialized(&mpi_init);