Support tensors with only column-wise data

[timmoon10]

Description

The quantized tensor infrastructure in TE 2.0 assumes that tensors have row-wise data available, both in the core C++ library and in the PyTorch extensions. This PR relaxes that assumption to support tensors wth only column-wise data. This allows us to avoid caching unnecessary data after the linear forward pass (we only need column-wise input for wgrad GEMM) and to reduce communication volume in MXFP8 tensor-parallel all-gathers. It is not quite perfectly optimized (tensor-parallel linear module still caches BF16 input tensor instead of column-wise MXFP8 tensor), but it is a reasonable step.

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

• Support tensors with only column-wise data in the core C++ library
• Support tensors with only column-wise data in the PyTorch framework
• Only cache column-wise data for input tensor in single-GPU Linear module, LayerNormLinear module, BasicLinear op
• Support MXFP8 all-gather with column-wise data

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

[timmoon10]

/te-ci core pytorch L0 L1