Move Mistral Medium recipe under canonical base model (#2150)

### What does this PR do?

Type of change: Bug fix

Moves the Mistral Medium 3.5 checkpoint-mirror PTQ recipe under the
canonical Hugging Face base model, `mistralai/Mistral-Medium-3.5-128B`.
Updates the recipe catalog, loader smoke-test path, and changelog.

### Usage

```bash
python examples/hf_ptq/hf_ptq.py \
    --model mistralai/Mistral-Medium-3.5-128B \
    --recipe huggingface/models/mistralai/Mistral-Medium-3.5-128B/ptq/nvfp4-max-calib
```

### Testing

- `uvx --from pre-commit pre-commit run --files CHANGELOG.rst
modelopt_recipes/huggingface/models/mistralai/Mistral-Medium-3.5-128B/ptq/nvfp4-max-calib.yaml
modelopt_recipes/ptq.md tests/unit/recipe/test_loader.py`
- Recipe documentation consistency checks (all passed)
- Focused pytest was unavailable locally because the checkout
environment lacks pytest/Torch; the dedicated pre-commit recipe
validator passed.

### Before your PR is "*Ready for review*"

- Is this change backward compatible?: ❌ The built-in recipe path
changes; the changelog documents the replacement path.
- If you copied code from any other sources or added a new PIP
dependency, did you follow guidance in `CONTRIBUTING.md`: N/A
- Did you write any new necessary tests?: ✅ Updated the built-in recipe
loader smoke test.
- Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?:
✅
- Did you get Claude approval on this PR?: N/A

### Additional Information

The recipe targets the vendor's canonical base checkpoint,
[`mistralai/Mistral-Medium-3.5-128B`](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B),
and reproduces NVIDIA's
[`Mistral-Medium-3.5-128B-NVFP4`](https://huggingface.co/nvidia/Mistral-Medium-3.5-128B-NVFP4)
quantization map.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added a Mistral Medium 3.5 128B post-training quantization recipe
supporting NVFP4 and FP8 configurations with max calibration.

* **Documentation**
* Updated checkpoint-mirror guidance to reflect the new model-based
recipe location.

* **Breaking Changes**
* Saved references to the previous recipe path must be updated to the
new Mistral Medium 3.5 recipe path.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: Chad Voegele <cvoegele@nvidia.com>

CHANGELOG.rst

@@ -22,6 +22,7 @@ Changelog
 
 **Backward Breaking Changes**
 
+- Move the Mistral Medium 3.5 checkpoint-mirror recipe from ``huggingface/models/nvidia/Mistral-Medium-3.5-128B-NVFP4/ptq/nvfp4-max-calib`` to ``huggingface/models/mistralai/Mistral-Medium-3.5-128B/ptq/nvfp4-max-calib``, keying it by the canonical Hugging Face base model. Update any saved ``--recipe`` paths to the new location.
 - Transformer Engine ``TEGroupedMLP`` (fused MoE experts) now uses **per-expert** weight quantization (one ``amax`` per expert) instead of a single shared ``amax``, so ModelOpt checkpoints containing quantized ``TEGroupedMLP`` modules saved before 0.47 are **not compatible** with 0.47. Re-run PTQ to regenerate compatible checkpoints.
 
 **Deprecations**

modelopt_recipes/ptq.md

@@ -235,7 +235,7 @@ that baseline. The deviations come in four kinds:
 | **Architecture-aware `quant_cfg`** | Per-sub-module format choices a single wildcard scheme can't express | `minimax_m3_vl`, `qwen3_5`, `qwen3_5_moe`, `vit`, `nemotron_llama` |
 | **Algorithm override** | Same numerics & scope, but the *calibration algorithm* is tweaked because the default breaks or regresses | `gemma`, `gemma4`, `mpt` |
 | **Extra exclusions** | Adds disabled-quantizer patterns so non-language branches stay full precision | `nemotron_vl`, `diffusion_gemma` |
-| **Checkpoint mirror** | A mixed-precision map reproducing one published checkpoint exactly | `models/nvidia/Nemotron-3-*`, `models/nvidia/Mistral-Medium-3.5-128B-NVFP4` |
+| **Checkpoint mirror** | A mixed-precision map reproducing one published checkpoint exactly | `models/nvidia/Nemotron-3-*`, `models/mistralai/Mistral-Medium-3.5-128B` |
 
 The numerics and standard exclusions are still inherited from `configs/`
 wherever possible — the model folder captures *only* the delta. Each `<task>/`
@@ -335,12 +335,12 @@ encoders (or the never-calibrated self-conditioning branch), regressing those
 modalities or crashing export. The extra patterns keep them in full precision;
 everything else matches the general recipe.
 
-### Checkpoint mirrors — `models/nvidia/<checkpoint>`
+### Checkpoint mirrors — `models/<org>/<checkpoint>`
 
 The `huggingface/models/` tier reproduces a **single published (or planned)
 checkpoint's** quant config verbatim:
 
-- **`Mistral-Medium-3.5-128B-NVFP4/ptq/nvfp4-max-calib`** mirrors
+- **`models/mistralai/Mistral-Medium-3.5-128B/ptq/nvfp4-max-calib`** mirrors
   `nvidia/Mistral-Medium-3.5-128B-NVFP4`: decoder MLP layers 4–86 use NVFP4
   W4A4, edge MLP layers 0–3 and 87 use FP8 W8A8, and all attention projections
   and the KV cache use FP8. It uses max calibration.

tests/unit/recipe/test_loader.py

@@ -167,7 +167,7 @@ _BUILTIN_PTQ_RECIPES = [
     "general/ptq/nvfp4_experts_only-kv_fp8",
     "general/ptq/nvfp4_experts_only-kv_fp8_cast",
     "general/ptq/nvfp4_experts_only-kv_fp8_layerwise",
-    "huggingface/models/nvidia/Mistral-Medium-3.5-128B-NVFP4/ptq/nvfp4-max-calib",
+    "huggingface/models/mistralai/Mistral-Medium-3.5-128B/ptq/nvfp4-max-calib",
     "general/ptq/nvfp4_mlp_only-kv_fp8",
     "general/ptq/nvfp4_mlp_only-novit-kv_fp8",
     "general/ptq/nvfp4_mlp_only-kv_fp8_cast",