Fail closed for non-affine NAX quantization

[PhilipJohnBasile]

Scope

MTPLX custom NAX and vk verification kernels consume affine quantization biases. This patch preserves stock MLX execution for every QuantizedLinear mode other than affine before any custom dispatch or affine-only tensor access.

Evidence

• Real nvfp4, mxfp4, and mxfp8 layers fall back to stock MLX with zero output difference.
• A simulated future non-affine 6-bit mode exercises the common guard.
• No custom verification kernel is called in those cases.
• Focused suite: 10 passed.
• Ruff, py_compile, and git diff --check passed.
• Mutation check: removing the guard reproduces the missing biases KeyError.
• Independent frozen-diff Sol review: APPROVE, no blockers.

This is a fail-closed compatibility fix, not a speed claim.

[PhilipJohnBasile]

Archiving this stacked 2.4-era component PR with its parent integration stack. It is not part of the current 2.9.1 mainline or upstream PR youssofal#336. The branch and commit remain preserved; any still-relevant affine-guard change should return later as a small rebase onto current upstream main.