Exploring Self-supervised Pre-trained ASR Models For Dysarthric and Elderly Speech Recognition