feat(kda): enable FlashKDA on SM100

[icavan]

Summary

• allow the SM90-derived FlashKDA CuTeDSL path on validated SM100 devices
• share one architecture guard between backend selection and the direct FlashKDA API
• keep SM103 disabled until it receives equivalent hardware validation
• document that this is the compatibility path, not the separate SM100-native modular implementation
• add architecture unit coverage and an SM100 public-dispatch correctness test

Validation

• pre-commit run --all-files
• GB200 SM100 on CUDA device 1: python -m pytest -q tests/test_kda_flashkda_arch.py tests/test_kda_flashkda_sm100.py -s
  • 6 passed
  • the SM100 test calls public cula.kda.kda_prefill, verifies auto intracard CP enters the FlashKDA CP driver, and compares output plus final state against FLA

Scope

This PR only changes architecture admission. The kernel pipeline remains the existing SM90-derived FlashKDA K1/K2 and intracard-CP implementation, which was compiled and executed successfully on SM100. It does not route FlashKDA to the newer SM100-native modular kernels.

[cherhh]

LGTM.