tiktoken's cl100k_base vocabulary, vendored so the tokenizer tests run with the network closed.

File:     9b5ad71b2ce5302211f9c61530b329a4922fc6a4
Upstream: https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken
SHA-256:  223921b76ee99bde995b7ff738513eef100fb51d18c93597a113bcffe865b2a7

tiktoken caches a downloaded vocabulary in TIKTOKEN_CACHE_DIR under the SHA-1 of its URL,
which is the file name above, and accepts a cached copy only when its SHA-256 equals the hash
tiktoken_ext/openai_public.py pins for that URL (the value above, as of tiktoken 0.14.0).
tests/conftest.py sets TIKTOKEN_CACHE_DIR to this directory for the suite only; nothing under
src/ reads it.

If a tiktoken release pins a different hash for this URL, tiktoken deletes this file and fetches
the vocabulary again, and the offline test fails. Replace the file with the new upstream bytes.

LICENSE beside this file is tiktoken's license (MIT), copied from the tiktoken 0.14.0 wheel.
