GOLLuM: Gaussian Process Optimized LLMs -- Reframing LLM Finetuning through Bayesian Optimization