Regularizing towards Causal Invariance: Linear Models with Proxies