Improved performance of mxfp8 cast kernels

[Oleg-Goncharov]

Description

Modified and tuned fused cast mxfp8 kernels for better performance.

Work is still in progress, and some further optimizations will be included in the next few days:

1. Caching the computed activations in the first pass to avoid recomputing them in the second pass (along the other dim)
2. Reducing the block size for CAST+DBIAS+DACT kernel specifically
3. Swapping the order of DBIAS computation

The performance comparison data will also be provided.

Fixes # (issue)

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

• Increased block size from 64x64 to 128x128
• Modified the scheme of how tensor elements are processed. ROWWISE: 32 elements per thread (i.e. one scaling factor)
• Added micro-optimizations (e.g. using MUL2 instructions; fusing MUL and CVT instructions directly in PTX)

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