Incorrect type annotation for encoder aggregation

The constructor of Seq2Seq defines aggregations as a `str`, but the forward method treats it as a sequence, and the jsonnet file splits from a comma-separated string to an array. I'm not sure if something has changed in allennlp, but for me, this results in an exception that it's expecting aggregations to be a string.

https://github.com/allenai/vampire/blob/cf5c2fe17caf73bbd0a470c48cba4d3ab4e847ff/vampire/modules/encoder.py#L76