[PyTorch] Document heterogeneous quantization recipes

[negvet]

Description

Document heterogeneous quantization recipes.
This is enabled after hybrid quantization support #2817 is merged

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR documents heterogeneous PyTorch quantization recipes and corrects the example’s wgrad provenance description.

• Adds conceptual guidance for role-aware quantizer factories, GEMM-direction mapping, hybrid quantizers, and high-precision slots.
• Adds a runnable heterogeneous quantization example and supporting diagrams.
• Expands API documentation for hybrid and identity quantizers and tensor types.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the previously reported wgrad provenance discrepancy is corrected and the current documentation matches the quantizer configuration and Linear backward path.

Important Files Changed

Filename	Overview
docs/examples/heterogeneous_quantization/pytorch_heterogeneous_quantization_example.py	The example now accurately describes wgrad as consuming the original BF16 columnwise input and output-gradient representations.
docs/features/low_precision_training/heterogeneous_quantization/heterogeneous_quantization.rst	Adds consistent documentation for role selection, GEMM-direction mapping, columnwise provenance, and high-precision quantization slots.
transformer_engine/common/recipe/init.py	Clarifies the CustomRecipe factory contract, including optional roles and the requirement to return a valid quantizer for every slot.
transformer_engine/pytorch/tensor/hybrid_tensor.py	Marks HybridQuantizer as experimental without changing runtime behavior.
transformer_engine/pytorch/tensor/identity_tensor.py	Expands IdentityTensor API documentation and marks IdentityQuantizer as experimental without changing runtime behavior.

_{Reviews (3): Last reviewed commit: "resolve comments" | Re-trigger Greptile}

[pggPL]

First observation: this part is done in very different style than the rest of the Features section.
Two most important things:

1. We use boxes for code.
2. It would be nice to have some pictures in .svg (they are quite easy to vibe code).

[pggPL]

First observation: when I start reading it there is a lot of defensive text about experimental API and some philosophical section Configuration readiness. I think it would be great to start, after short introduction with short example what we can do - using more than one recipe with the picture with some code snippet. Then we can elaborate on the API in more detail.

[pggPL]

I think this is too long - we can have something before #START_FINE_GRAINED_QUANTIZATION_EXAMPLE or in different files like model building and require_supported_hardware() . BTW see how we deal with support hardware in boxes in the other recipes docs.

[pggPL]

Note that we can split this: short docs in Features and tutorial in Tutorials like we have done with gemm profiling tutorial.

[negvet]

@pggPL thanks for the comments, I applied all your suggestions