cif-based collaborative decoding for end-to-end contextual speech recognition