Computational and sensitivity aspects of eigenvalue-based methods for the large-scale trust-region subproblem