Harden Python profiler command handling

[fallintoplace]

Summary

• Fix the CUTLASS 3.x profiler layout fallback so row-major output kernels preserve the A/B layout tags when selecting a column-major-output profiler kernel.
• Run the profiler with list-form subprocess.run instead of building a shell command string.
• Parse profiler metrics with checked regex matches and raise explicit errors for parse or metric mismatches.
• Use GemmCoord properties consistently when passing problem sizes and computing expected bytes/FLOPs.

Testing

• PYTHONPATH=python /Users/hoangvu/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3 test/python/cutlass/profiler.py
• git diff --check