Learning Joint Nonlinear Effects from Single-variable Interventions in the Presence of Hidden Confounders