Learning Theory for Kernel Bilevel Optimization