Unidentified and Confounded? Understanding Two-Tower Models for Unbiased Learning to Rank