Invariance in Policy Optimisation and Partial Identifiability in Reward Learning