Random seeding overhaul

Right now setting the random_seed parameter in the audit config does nothing, the only part of the code base that seeds everything is using the optuna seed at the start of optuna trials.
Ideally it should seed everything when:
- Starting an audit
- Starting an attack (preferred for reproducability when some results are already run)

Also, the current seed_everything function used by optuna is problematic in two ways:
- Deterministic cudnn is not necessary for most models and usually slows down training/inference by some 30% (in some cases 100x slower)
- Setting cudnn in benchmarking mode has nothing to do with seeding, but is generally good for performance