LNUCB-TA: Linear-nonlinear Hybrid Bandit Learning with Temporal Attention