Fine-tuning LLMs with variational Bayesian last layer for high-dimensional Bayesian optimization