Transformer-based Models for Long-Form Document Matching: Challenges and Empirical Analysis