Remove a buggy/redundant reset

Author: nanz-nv

megatron/core/full_cuda_graph.py

@@ -189,7 +189,6 @@ def __call__(self, *args, **kwargs):
             torch.cuda.synchronize()
             torch.distributed.barrier()
             logger.info(f'CUDA graph capture done!!!')
-        paged_stash_reset(enabled=self.moe_paged_stash and training)
         if FullCudaGraphWrapper.cuda_graph[training_str] is None:
             FullCudaGraphWrapper.result[training_str] = self.forward_backward_func(*args, **kwargs)
         else: