Change the save type of the inference output

[SamuelBrand1]

At the moment the pipeline does a produce_or_load call from DrWatson.jl. This defaults to saving the produced dictionary of output from the inference as a JLD2 file.

JLD2 is a decent HDF5 format, so compresses nicely. However, it saves information about the types of the variables being serialised... and this means that if the underlying EpiAware types change due to package dev then deserialisation becomes difficult.

Its probably more robust to output to parquet.