Shared versus reservoir-specific policy architectures in multi-agent reinforcement learning for physically heterogeneous cascade reservoir operation