Non-markovian reward modelling from trajectory labels via interpretable multiple instance learning