Squeezing Lemons with Hammers: An Evaluation of AutoML and Tabular Deep Learning for Data-Scarce Classification Applications