Mixed-precision Orthogonalization Scheme and Adaptive Step Size for CA-GMRES on GPUs