Fix(tokenizing): Use multi-core

[NanoCode012]

This improves speed for tokenizer and adds a progress bar.

Benchmark:

dataset	old	new
teknium/GPT4-LLM-Cleaned	~70s	< 10s
WizardLM/WizardLM_evol_instruct_V2_196k	~5 min	~55s

Note: just internally counting, nothing strict.

Future PR for more speed can:

• use batching. (requires changing class to not use iter)
• fix Dataset.from_list and shuffle (do we need to save to a list? why not save to a dataset directly to cut off this part)
• ConstantIterList can be improved

Full credit to: neverendingtoast

[winglian]

🔥

[NanoCode012]

One note: this removes the error if dataset is empty. We can simply just add into it the data.py file instead.