`contextualized_embed` can return array of `nan`

There are certain inputs which deterministically get an array of `nan` for the embedding of the last chunk when `encoding_format` is not set (I assume because it's then defaulting to b64 and something is going wrong). Can provide a sample if you email me!