Learning Active Camera for Multi-Object Navigation