QOL changes for generations

[maxmatical]

1. append task_name for multiple tasks
2. fix n_tasks logic from dataset index out of bounds

Next 2 changes related to #58
4. save intermediate generations with --save_every_k_samples
5. resume generation from intermediate generations with --load_generations_intermediate_paths

Tested with HumanEval with saving every 50 samples + loading from intermediate generations:

len(intermediate_generations) = 150
should be generating 14 new samples for new_generations
curr_sample_idx = 150
number of problems for this task is 14
len(dataloader)= 14
len(code_gens) = 14

len(new_generations) = 14
len(generations) after concatenating = 164

Verified:

• loading form intermediate generations generates same output as with --limit_start 150 on HumanEval
• Saved generations match final generations
• Eval metrics unchanged

5. (minor) add some typing + linting

[RaymondLi0]

Nice work, thank you!
I am wondering whether there would be an issue with several restarts (see comment)

[RaymondLi0]

if there are multiple saving steps, I think we'll add the same generations several times into intermediate_generations.
Also this list is also extended in evaluator.py

bigcode-evaluation-harness/bigcode_eval/evaluator.py

Line 82 in 0754793

generations.extend(new_generations)

[maxmatical]

you're right, made 2 changes in the new commit

1. instead of extending on intermediate_generations when saving which end up duplicating new generations, extend on a deepcopy instead which prevents duplications
2. instead of extending at evaluator, return intermediate_generations.extend(code_gens) in complete_code, which makes a bit more sense with this new logic. since we never mutate intermediate_generations when saving, this will return the non-duplicated generations

[RaymondLi0]

One small suggestion. Then good to merge! Thank you