docs: add datatype and hardware support matrix

[CodersAcademy006]

Problem

There is no single place in the docs that states which low-precision formats
are supported on which GPU architecture. The information is spread across the
README, inline compute-capability checks, and test skip conditions. Several
issues ask for exactly this (#2512, #2304), and related threads keep hitting the
same confusion about SM120 support (#2668, #2255).

Solution

Add docs/support_matrix.rst, a page that documents format support keyed by
compute capability, since that is the value Transformer Engine checks at runtime.
It covers BF16, FP8 per-tensor, FP8 block scaling, MXFP8, and NVFP4, plus the
default recipe selected per architecture.

The tables are derived directly from the runtime capability checks so docs and
behavior stay aligned:

• _compute_fp8_support, _compute_mxfp8_support, _compute_nvfp4_support,
  _compute_fp8_block_scaling_support, and get_default_fp8_recipe in
  transformer_engine/pytorch/quantization.py
• the BF16 check in transformer_engine/pytorch/utils.py

Two nuances are called out explicitly because they are common sources of
confusion:

• MXFP8 is not currently supported on compute capability 12.0 and higher; it is
  limited to 10.0 through 10.x.
• NVFP4 is reported as available on 10.0 and higher, but the default recipe uses
  stochastic rounding whose FP4 conversion instructions are architecture specific
  to 10.0 and 10.3. On 12.0 the default recipe raises an architecture-specific
  error and round-to-nearest is required.

The page is linked under the Getting Started section in docs/index.rst.

Testing

Validated the reStructuredText structure with docutils. I was not able to run a
full Sphinx build locally because that requires a compiled Transformer Engine,
which needs a CUDA GPU I do not have access to. A reviewer running the docs build
in CI can confirm the page renders as expected.

Limitations

This page documents the current state of the runtime capability checks. If those
checks change, the tables need to be updated to match. A note in the page states
this so the two stay in sync.

Closes #2512
Closes #2304
Refs #2668
Refs #2255

[greptile-apps]

Greptile Summary

This PR adds a documentation page for low-precision datatype support. The main changes are:

• A new support matrix page for GPU compute capabilities.
• Tables for BF16, FP8, FP8 block scaling, MXFP8, NVFP4, and default recipes.
• A main docs index entry linking the new page.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

Important Files Changed

Filename	Overview
docs/index.rst	Adds the new support matrix page to the main documentation toctree.
docs/support_matrix.rst	Adds datatype support tables and caveats that match the documented runtime capability checks.

_{Reviews (1): Last reviewed commit: "docs: add datatype and hardware support ..." | Re-trigger Greptile}