fix: use PreTrainedTokenizerFast wrapper for MLXLM in xgrammar and llguidance backends

[octo-patch]

Fixes the same class of bug reported in #1847, which noted that code in `mlxlm.py` was passing the raw `tokenizers.Tokenizer` backend instead of the `PreTrainedTokenizerFast` wrapper to downstream APIs.

Problem

Two backends had the same incorrect pattern when handling MLXLM models:

`xgrammar.py` (line 129):

# Before (wrong):
tokenizer = model.mlx_tokenizer._tokenizer  # raw tokenizers.Tokenizer
xgr.TokenizerInfo.from_huggingface(tokenizer, ...)

`llguidance.py` (line 226-227):

# Before (wrong):
llguidance.hf.from_tokenizer(model.mlx_tokenizer._tokenizer)  # raw backend

Both `xgr.TokenizerInfo.from_huggingface` and `llguidance.hf.from_tokenizer` expect a `PreTrainedTokenizerFast` wrapper (which has `eos_token_id`, `eos_token`, `all_special_tokens`, etc.), not the raw `tokenizers.Tokenizer` backend.

With modern `transformers` / `tokenizers` >= 0.25, the raw backend no longer exposes these attributes, causing `AttributeError` when MLXLM is used with the xgrammar or llguidance backends.

Solution

Replace `model.mlx_tokenizer._tokenizer` with `model.mlx_tokenizer` in both backends. This is consistent with:

• The outlines_core backend, which already correctly uses `model.mlx_tokenizer` (line 200 of `outlines_core.py`)
• The Transformers model path in both backends, which uses `model.hf_tokenizer` (the `PreTrainedTokenizerFast` wrapper)

Testing

Existing MLX integration tests in `tests/backends/test_xgrammar.py` and `tests/backends/test_llguidance.py` cover these code paths when run on Apple Silicon (`HAS_MLX=True`). The backends MLXLM branches are marked `# pragma: no cover` as they require Apple Silicon to execute.