Contextual morphologically-guided tokenization for Latin encoder models