Multiply robust causal inference of the restricted mean survival time difference