Make TiktokenTokenizerWrapper picklable

[irenedea]

Makes TiktokenTokenizerWrapper serializable so we can use it with multi-processing.

Tests

1. Finetune run with tiktoken tokenizer: test-pickle-tiktoken-ohOd0p
2. Works with packing
3. pickle code:

from llmfoundry.tokenizers import TiktokenTokenizerWrapper
t = TiktokenTokenizerWrapper('gpt2')
import pickle
pickle.dumps(t) # works

(Previously, the above code would raise an error: TypeError: cannot pickle 'builtins.CoreBPE' object)

[samhavens]

Do you know why we are getting /home/runner/work/llm-foundry/llm-foundry/build/lib/llmfoundry/tokenizers/tiktoken.py:59:43 - error: "Encoding" is not a known member of module (reportGeneralTypeIssues) in the Code Quality checks? I was getting this too. Is the CI environment not getting a full install of the tokenizer or something?

[irenedea]

@samhavens I think we might need to add the openai deps to the .github workflow file. Will try it!