Robust Invariant Representation Learning by Distribution Extrapolation