Fix Qwen3_5 mixed-bit per-tensor quantization keys not remapped through sanitize

[ivaniguarans]

Problem

Loading an Qwen3_5ForConditionalGeneration checkpoint with non-uniform per-tensor config["quantization"] overrides — notably Intel's AutoRound mixed-bit MLX builds — fails on main with:

ValueError: Expected shape (10240, 640) but received shape (10240, 960)
for parameter language_model.model.layers.0.linear_attn.in_proj_qkv.weight

Reproduced against Intel/Qwen3.6-27B-4.5b-mlx-AutoRound (459 per-tensor entries, 128 of which are real 6-bit overrides on top of a 4-bit global).

Root cause

class_predicate in mlx_lm/utils.py looks up paths in the post-sanitize namespace, but config["quantization"] keys are stored in the pre-sanitize namespace as they were written to config.json. For Qwen3.5_VL, Model.sanitize in mlx_lm/models/qwen3_5.py rewrites model.language_model.* → language_model.model.* (and drops vision_tower.* / model.visual.* entries), but the quantization config is never remapped to match. Every per-tensor override misses the lookup, falls back to the global bit-width, and the tensor is allocated at the wrong size — failing the shape check on load.

Fix

Add a per-model sanitize_quantization(quant_config) method on Qwen3_5ForConditionalGeneration that mirrors the existing sanitize(weights) key transformation, and wire it via a hasattr-discovered opt-in call at the _quantize call site in mlx_lm/utils.py — the same shape as the existing sanitize invocation a few lines above. Qwen3_5MoeForConditionalGeneration inherits the method through its class Model(Qwen3_5Model) subclassing, so MoE checkpoints pick up the prefix remap without a separate implementation (MoE-specific expert-key remaps such as experts.gate_up_proj → switch_mlp.gate_proj are not in scope of this change — no MoE mixed-bit checkpoint with per-tensor expert overrides has been reported yet).

Models without the method are unaffected. The hook is scoped to the top-level config["quantization"] branch; the legacy quantization_config path (AWQ/GPTQ/mxfp4) carries flat global configs without per-tensor overrides and does not exercise this bug.

Credit to @ivanfioravanti for the local-fix observation in #1214's comments.

Testing

• Unit tests (TestQwen3_5SanitizeQuantization in tests/test_utils.py): one table-driven test covering all five key shapes the method handles (globals passthrough, vision_tower / model.visual drops, model.language_model.* rewrite, language_model.* passthrough, bare-key prefix), plus a non-mutation test confirming the input dict is left intact.
• Loader-path integration test (TestUtils.test_load_model_qwen3_5_with_unsanitized_per_tensor_quantization): mirrors the existing test_load_model_gemma4_with_per_layer_projection_quantization pattern. Builds a tiny synthetic Qwen3.5 model, hand-edits config["quantization"] with an un-sanitized per-tensor override key, saves and reloads via load_model, and asserts the rewritten key appears in loaded_config["quantization"] (the un-sanitized form is gone), then runs a forward pass as a smoke test that load_weights doesn't raise on the rewritten config.
• Real-checkpoint verification on M5 Max (128 GB) against Intel/Qwen3.6-27B-4.5b-mlx-AutoRound:
  • Without this patch: ValueError as quoted above.
  • With this patch: load succeeds in 1.3 s; generate(prompt="The capital of France is", max_tokens=8) returns "Paris.\n\n<think>\nHere's a".
• Full python -m unittest discover tests/ shows no regressions (12 pre-existing failures, all verified to fail identically on upstream/main).

Fixes #1214.