MCP: Learning Composable Hierarchical Control with Multiplicative Compositional Policies