Resolve tokenizer from local cache in offline mode

[sarathfrancis90]

Fixes #248

Bug

With HF_HUB_OFFLINE=1 and a model fully present in the local cache, download_tokenizer_from_hf_hub(repo_id, revision="main") crashes instead of resolving the tokenizer from the cache. This is hit in practice via vLLM / transformers v5, which route any model shipping tekken.json to MistralCommonBackend and pass revision="main".

Cause

Two issues combine in tokens/tokenizers/utils.py:

1. In offline mode huggingface_hub raises OfflineModeIsEnabled, which is a builtin ConnectionError subclass, not a requests.* error. The except (requests.ConnectionError, requests.HTTPError, requests.Timeout) around list_repo_files therefore misses it and the existing local-files fallback is skipped.
2. list_local_hf_repo_files only consulted refs/<DEFAULT_REVISION> when revision was None. An explicit branch/tag like "main" was looked up as a literal snapshots/main directory, which never exists (snapshots are keyed by commit hash), so it returned [].

Fix

• Add huggingface_hub.errors.OfflineModeIsEnabled to the caught exceptions so offline mode reaches the same local-files fallback as a dropped connection (force_download still re-raises, unchanged).
• Resolve any branch/tag revision to its commit hash via refs/<revision> before reading the snapshot directory.

Testing

• Added a revision="main" case to test_list_local_hf_repo_files and a new test_download_tokenizer_from_hf_hub_offline_mode; both fail before the change and pass after.
• pytest tests/test_utils.py (23 passing), ruff check, ruff format --check, and mypy src are all green.