Do Explicit Alignments Robustly Improve Multilingual Encoders?