On the Limitations of Cross-lingual Encoders as Exposed by Reference-Free Machine Translation Evaluation