Add cosine similarity tests and allow schemad data

[RobinL]

This allows tests to be specified as a pyarrow table which in turn means that a stricter schema can be imposed on the incoming data.

In this case, using DuckDB, the cosine similarity function can only be used if the data is a fixed length array. When passing a python list to pyarrow, by default, this is treated as a variable length array and hence the old framework was not quite flexible enough to accomodat e