How Toxicity Classifiers and Large Language Models Respond to Ableism