fix tabpfn categorical handling

[alperkent-cmi]

tabpfn currently disregards the input variable schema and just interprets all columns with <30 unique values to be categorical. fix this so it respects the input schema.