Upgrade transformers to 5.x (drops conftest mistral monkey-patch)

[voorhs]

Background

tests/conftest.py currently monkey-patches transformers.PreTrainedTokenizerBase._patch_mistral_regex to a no-op as a workaround for a confirmed transformers bug: on every tokenizer load with vocab_size > 100000 (e.g. our default intfloat/multilingual-e5-*), the method calls huggingface_hub.model_info() unconditionally. On CI that hammers the 1000-req/5-min HF rate limit and 429s.

Upstream issues / PRs documenting the bug:

• AutoTokenizer.from_pretrained calls model_info() unconditionally in _patch_mistral_regex, breaks HF_HUB_OFFLINE mode huggingface/transformers#44843
• 4.57.3 issue with AutoTokenizer.from_pretrained when HF_HUB_OFFLINE=1 huggingface/transformers#43200
• Transformers is trying to call home despite local_files_only=True huggingface/transformers#45545
• [fix] Always early return for non-Mistral models in _patch_mistral_regex huggingface/transformers#45444
• Fix unintended Hub metadata calls from _patch_mistral_regex huggingface/transformers#43603

The #45444 merge commit message explicitly says:

faulty mistral tokenizers saved with version 4.57.{3,4,5,6} will load incorrectly... use 5.0.0 as 5.0.0rc0 introduced the fix... Exclusively check for mistral BEFORE v5

So:

• transformers 4.57.x has the bug — we're on 4.57.6, confirmed buggy
• transformers 5.0.0+ has the fix — latest is 5.10.2 (released yesterday)
• The fix was not backported to 4.x — it's "v5-only" by design

Task

Upgrade transformers to >=5.0 in pyproject.toml, then delete _disable_transformers_mistral_regex_patch() from tests/conftest.py.

Risk

transformers 5.0 is a major version bump (Mistral 4 support, removed deprecated APIs, etc). Likely surfaces breaking changes elsewhere in the codebase that would need addressing separately.