Escaping Saddle-Points Faster under Interpolation-like Conditions