[GPU] Using a cpu impl type for NMS in a dynamic case (#21648)

Author: Lyamin-Roman

src/plugins/intel_gpu/src/graph/layout_optimizer.cpp

@@ -1493,7 +1493,7 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
             } else {
                 const auto scores_layout = nms_node.input_scores().get_output_layout();
                 if (scores_layout.is_dynamic()) {
-                    preferred_impl = impl_types::ocl;
+                    preferred_impl = impl_types::cpu;
                 } else {
                     const size_t kBatchNum = scores_layout.batch();
                     const size_t kClassNum = scores_layout.feature();

src/plugins/intel_gpu/src/graph/primitive_inst.cpp

@@ -1404,7 +1404,8 @@ memory::ptr primitive_inst::allocate_output(engine& _engine,
 
     // For outputs, cpu prim we want to have lockable alloc type
     // Also if the successor of a node is an cpu, then memory needs to be lockable.
-    bool is_cpu = _node.get_selected_impl() ? _node.get_selected_impl()->is_cpu() : false;
+    bool is_cpu = _node.get_selected_impl() ? _node.get_selected_impl()->is_cpu() :
+                                              _node.get_preferred_impl_type() == impl_types::cpu;
     auto use_lockable_memory =
         is_output_buffer || is_cpu ||
         has_any_cpu_user_not_shape_of(_node.get_users()) ||