[WIP] Eval scripts, harmfulness results, fix to eval_harmfulness, readme

[Adamliu1]

This PR is for scripts to evaluate tasks and evaluate harmfulness, but it includes some of my previous tasks regarding eval-harmfulness that I didn't push to the main.

Reference scripts:

• For eval harmfulness

run_experiment_multiple_seed_cluster.sh
run_experiment_multiple_seed.sh
eval_multi_seed_cluter_submission.sh 

NOTE: Seeding still seems to have a bit of a problem in the code; I noticed I got the same results from all models after using the reproducibility method from ADL CW. But currently, this is a low priority.

• For eval task framework (from Andrzej)

eval_results.sh
run_experiment_raw_model.sh 
run_experiment.sh schedule_experiment.qsub.sh

WIP

• For a script that evaluates all unlearned models (from experiments) and runs eval task framework and harmfulness evaluation.

run_evals.sh 
run_task_eval_exp.sh (small modification to Andrzej's script)
run_harmfulness_eval_exp.sh (missing)

[Willmish]

leaving open for now to have track of the experiment results done by @Adamliu1