Remove inference batching

[michaelharrisonmai]

Removed batching, swapped some logic of the dataloader / ThreadPoolExecutor / writer in inference.py, refactored the _run methods.

Current flow is: when run is called, load partial results if appropriate, then kick off ThreadPoolExecutor, which calls "_run_single" for each element of the dataloader and then appends the results to the file (note that appending vs writing is new -- let me know if this is not preferred for any reason). _run_single is a combination of the previous _run and _run_par: it checks for previous results, checks for rate limiting, and finally calls the model's generate().