tests: make the MLX boundary guard catch mlx_lm imports

[VKVstudio]

What

tests/test_repo_hygiene.py::test_mlx_imports_stay_inside_the_backend enforces the rule stated in backends/__init__.py: MLX packages are imported only under edge0/backends/mlx/. Its regex requires a space or a dot right after mlx, so from mlx_lm.models... import ... never matches and the guard silently passes framework code that imports mlx_lm directly.

This PR tightens the pattern to ^\s*(?:from|import)\s+mlx(?:_lm)?\b.

What it finds

The stricter guard immediately flags src/edge0/streaming/layer.py:49:

from mlx_lm.models.switch_layers import _gather_sort, _scatter_unsort

Rather than turn CI red on an existing file, that path is listed in an explicit _MLX_IMPORT_KNOWN_VIOLATIONS set with a note on what should move into the framework. Remove the entry once the two helpers no longer come from mlx_lm. I'm happy to send that follow-up: they are a few lines of argsort/gather that can live next to moe/routing.py, which would also make the streaming layer importable on a non-MLX backend.

If you would rather have the guard fail loudly instead, drop the allowlist — the regex change alone is the fix.

Checks

• pytest tests/test_repo_hygiene.py — 4 passed. The file imports nothing from edge0 or mlx, so this runs on any platform, same as the existing Linux hygiene job.
• The tightened regex matches import mlx, import mlx.core as mx, from mlx_lm import x, from mlx_lm.models.base import ..., and does not match from edge0.backends.mlx._impl import ... or a hypothetical mlxfoo package.

Context

I'm working on a CUDA/PyTorch backend for the reserved backends/cuda/ slot (https://github.com/smartbulldog-pro/edge0-cuda/tree/cuda-backend). This import is the one thing that keeps streaming/layer.py from importing on a non-MLX backend, which is how I ran into the guard.