Make sure fenced tensor could not reuse other tensor. (#2561)

Fix random error caused by this.

onnxruntime/core/framework/allocation_planner.cc

@@ -325,6 +325,7 @@ class PlannerImpl {
     auto p_required_buffer_shape = context_.GetShape(output_arg);
     if (nullptr == p_required_buffer_shape) return false;
     auto& required_memory_info = AllocPlan(output_arg.Name()).location;
+    if (HasFence(&output_arg)) return false;
 
     for (auto it = freelist_.begin(); it != freelist_.end(); ++it) {
       size_t reusable = static_cast<size_t>(it->ml_value);