Add no-op execution mode to support profiling lazy trace overhead (#69513)
- use python binding from lazy_bench script with new mode
--run_tracing_execute_noops
which works together with filter args to select a model and
repeatedly trace it without any backend execution
