Scenario decomposition of risk-averse multistage stochastic programming problems