Robust estimation of heterogeneous treatment effects in randomized trials leveraging external data