fix(mlx): tolerate invalid UTF-8 VLM tokens

[accelizero]

Motivation

Older mlx-vlm releases strictly decode the pending byte-level BPE buffer. A generated byte-fallback token followed by a space-prefixed token can raise UnicodeDecodeError and abort the full MinerU VLM task.

Modification

• Backport the upstream replacement-decoding behavior for older mlx-vlm releases at MLX engine initialization.
• Leave releases that already use errors="replace" unchanged.
• Add focused regression coverage for both paths.

Validation

• Focused Ruff check passes for the new compatibility module and regression test.
• Regression tests pass: 2 passed.
• Verified a one-page Hybrid MLX parse on macOS with MinerU 3.4.4, mlx-vlm 0.3.9, mlx 0.31.1, and transformers 4.57.6.

BC-breaking

No. The patch only prevents an MLX VLM generation crash; non-MLX backends and already-fixed mlx-vlm versions are unchanged.

This branch is based on the locally tested Git commit. The earlier web-editor PRs were closed because their new files had incorrect nested paths.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[accelizero]

I have read the CLA Document and I hereby sign the CLA

[accelizero]

recheck