Getters / setters for the codec of the tables energyRt writes to disk: a
scenario's parameter and variable stores, and the model / repository /
dataset stores. "feather" (Arrow IPC) and "parquet" both compress and
both support lazy, filtered reads through open_dataset(); parquet is
smaller, feather reads faster. "csv" is uncompressed and loses column
types on the round trip.
The format is recorded in the store's manifest and is authoritative when an
existing store is rewritten, so changing this option never mixes codecs
inside one store. The content-addressed stores (save_model(),
save_repository(), save_dataset()) pin parquet regardless, because their
identity is a content hash that must be reproducible.
For the files handed to a solver, see get_exchange_format().
