Add regression test for offline tokenizer loading (fixes #43200)

[Anri-Lombard]

Summary

This PR adds a regression test for issue #43200 where AutoTokenizer.from_pretrained() failed in offline mode (HF_HUB_OFFLINE=1) even when the model was cached locally.

Good news: The underlying bug was already fixed on main as part of the tokenizer refactoring. The _patch_mistral_regex function now correctly checks is_offline_mode() and sets is_local = True to prevent network calls (see tokenization_utils_tokenizers.py lines 1177-1178).

This PR adds a test to prevent regression.

Root Cause (for reference)

In v4.57.3, commit d3ee5e8 added _patch_mistral_regex which called model_info() without handling offline mode:

def is_base_mistral(model_id: str) -> bool:
    model = model_info(model_id)  # No error handling for offline mode
    ...

if _is_local or is_base_mistral(pretrained_model_name_or_path):

When HF_HUB_OFFLINE=1 and using a model ID (not local path), is_base_mistral() was called and model_info() raised OfflineModeIsEnabled.

The Fix (already on main)

The refactored code in tokenization_utils_tokenizers.py now includes:

if is_offline_mode():
    is_local = True

This ensures is_base_mistral() is never called in offline mode due to short-circuit evaluation.

Test Plan

• New test test_offline_tokenizer passes
• All existing TestOffline tests pass
• Verified offline tokenizer loading works manually