Use threading to speed up data iteration in tokenization

[mbsabath]

Currently in tokenizer, examples are loaded into the queue one example at a time in a single threaded manner. we should be able to have multiple concurrent requests loading, assuming that hugging face iterable datasets are thread safe

[mbsabath]

In experimentation, we've found that huggingface streaming dataset iterables do not support asynchronous iteration, this may require deeper work in the future