[Refactor] Verbosity rework

[lintangsutawika]

Make the args.verbosity work as initially intended across the whole repo.

The command

lm-eval --model hf --model_args "pretrained=EleutherAI/pythia-2.8b" --task mmlu --limit 10 --verbosity WARNING

now suppresses DEBUG level logs in the lm_eval/tasks/__init__.py

I'll inspect and see if other logs in lm_eval should be set to DEBUG instead of INFO.

cc: @haileyschoelkopf

[lintangsutawika]

@haileyschoelkopf

1. Added initialize_task to handle all task initialization so that it doesn't do it when lm_eval.tasks or other modules are imported.
2. Imports are moved back to the top
3. eval_logger is part of utils.py

[haileyschoelkopf]

this looks good, just some minor comments about style, and about making sure that evaluate() functions work outside of running main.py.

I think we'll also have to make sure the scripts/ directories call initialize_tasks() as well.