Towards Effective And Efficient Open Speech Foundation Models