Attribute-efficient learning of decision lists and linear threshold functions under unconcentrated distributions