Options as responses: Grounding behavioural hierarchies in multi-agent RL