Stale MXFP4 training entry in workflow documentation

Documentation mismatch

As a torchao user comparing release/0.16.0 and release/0.18.0, I noticed that the MXFP4 table in docs/source/workflows/training.md still contains descriptions from the former general MX training workflow:

• MXFP4 training for torch.nn.Linear is shown as available.
• MXFP4 QAT is shown as planned.
• Enabling MXFP4 QAT is listed as a planned improvement.

These entries do not match the current source:

• In 0.16, MXLinearConfig exposed the MXFP4_EMULATED and MXFP4_CUTLASS training recipes.
• #3968 removed MXLinearConfig, MXLinear, and the corresponding training tests.
• The current low-precision MX training configuration and linear path are MXFP8-specific.
• The current MXFP4 workflows are inference and QAT. MXFP4 QAT is implemented in torchao.prototype.qat.mx and covered by forward/backward tests.

The same stale table is present in release/0.18.0 and the current main branch.

Proposed documentation correction

• Remove the obsolete MXFP4 quantized-training row.
• Update MXFP4 QAT from planned to its current prototype status.
• Remove the obsolete planned-improvement item about enabling MXFP4 QAT.

This is a documentation-only correction and does not propose adding or removing any runtime functionality.

Open docs/source/workflows/training.md and compare the MXFP4 table with the current MX training configuration and the MXFP4 implementation in torchao.prototype.qat.mx. Remove the obsolete quantized-training and planned-improvement entries, and mark MXFP4 QAT as implemented prototype functionality consistent with the referenced forward/backward tests.