[TensileLite] Fix MX FP4 scale data initialization & argument scoping

[archana-ramalingam]

Motivation

Fix MXFP4 scale data initialization and kernel argument scoping

Technical Details

• DataInitialization.cpp: Extend initializeCPUInputs skip guard to include MXSA/MXSB — the general tensor loop was overwriting correlated scale data from initializeMXDataForFP4 with independent random UE8M0 values, causing intermittent single-element validation failures from sub-ULP hardware/software dequantization differences UE8M0→float32
• ContractionSolution.cpp: Gate scaleA/scaleB argument appending on useScaleAB only — MX problems without UseScaleAB were incorrectly passing scale pointers to kernels that don't expect them
• Signature.py: Fix indentation of userArgumentsInfo size increments for scaleA/B/C/D, bias, and E tensor — sizes were unconditionally accumulated outside their feature if blocks, producing incorrect kernel argument layouts

Test Plan

TBD

Test Result

TBD

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.