support MXFP8 (#21146)

[metascroy]

Summary:

Add MXFP8 block-scaled float quantization support to the MLX backend, alongside
the existing NVFP4 path, and unify the MLX safetensors checkpoint loader so it can
load block-scaled (NVFP4 / MXFP8) checkpoints directly.

• New ExportableMXFP8Tensor (extension/llm/export/mxfp8.py): a torchao tensor
  subclass carrying MLX-native packed weights (uint32 qdata, uint8 E8M0 block
  scales, group_size 32, 8-bit FP8), with a registered torchao::dequantize_mxfp8
  custom op so the dequant survives torch.export, and a payload-preserving .to()
  so activation-dtype casts do not dequantize it.
• NVFP4 (extension/llm/export/nvfp4.py): add the same payload-preserving .to()
  so block-scaled weights survive model.to(activation_dtype) instead of being
  dequantized.
• MLX backend lowering (backends/mlx/patterns.py, backends/mlx/builder/op_helpers.py):
  add parse_dequant_mxfp8_node and MXFP8 linear/embedding pattern handlers that
  emit the MLX mode="mxfp8" quantized matmul / gather (group_size 32, bits 8).
  The bundled MLX runtime already accepts this mode, so no C++/runtime changes.
• Fusion (extension/llm/export/quant/convert.py): extend fuse_along_output and
  maybe_cast to handle the NVFP4 / MXFP8 carrier tensors (the per-tensor scale
  is treated as a shared, assert-equal field rather than concatenated).
• Checkpoint loader: unify the MLX reader to resolve each weight's quant mode
  per-tensor from config.json (a default mode plus per-module overrides) and
  build the matching representation -- affine -> torchao intx, nvfp4 ->
  ExportableNVFP4Tensor, mxfp8 -> ExportableMXFP8Tensor -- so mixed-format
  checkpoints load correctly from a single path.

Reviewed By: Gasoonjia

Differential Revision: D113113318

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21146

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 27 Pending

As of commit 1a5735a with merge base 96e87e3 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[meta-codesync]

@metascroy has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113113318.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.