On the capabilities of LLMs for classifying and segmenting time series of fruit picking motions into primitive actions