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