specialization_run_init used prepare_exact_args_frame which heap-allocates
a FrameObject with 4-5 atomic refcount bumps per call. Switch to
invoke_prepared_exact_args which uses InterpreterFrame::new_on_datastack
with zero refcount bumps.
CallAllocAndEnterInit already guards against tracing (eval_frame_active)
and recursion limits before reaching this path.
Assisted-by: Claude