--learning mode

[karmacoma-eth]

we need better ergonomics for learning the best sequence of solvers for a particular set of queries.

The current approach:

# run jsi on every file, producing csv timings
for smt_file in path/to/smt_files/*.smt2 ; do jsi --csv --full-run --timeout 1m "$smt_file" ; done

# ingest csv and display stats about best solvers
python utils/csv-stats.py path/to/smt_files/

We could instead have something like:

jsi --learning [--sequence a,b,c] path/to/smt_files/

This could optionally start with sequence a,b,c, but use heuristics to prioritize solvers that have worked best on previous queries