Update SM90 FlashKDA benchmark results

[icavan]

What changed

• remove the two obsolete benchmark Markdown generator scripts
• replace the legacy CUTLASS C++ SM90 numbers with current CuTe DSL FlashKDA measurements
• add intracard CP-auto vs CP-off results for long and ragged sequence shapes
• update README highlights and point reproduction instructions at the benchmark scripts directly

Why

SM90 now defaults to the CuTe DSL FlashKDA backend, but the published Hopper data still described the older C++ implementation. This updates the documentation to reflect the current backend and includes the low-head, long-sequence cases where intracard CP is intended to help.

Performance

Measured on an H200 141GB / SM90 GPU with 132 SMs, PyTorch 2.9.1+cu129, Triton 3.5.1, and FLA v0.5.0:

• FlashKDA vs FLA: 2.39x average across 28 fixed/variable-length configs; 1.38–7.80x range
• intracard CP engaged in 28/32 configs: 4.11x geometric mean over serial FlashKDA; 1.93–7.21x range

Validation

• pre-commit run --all-files
• python -m pytest tests/test_kda_sm90_prefill_vs_fla.py tests/test_kda_sm90_intracard_cp.py -q
  • 28 passed, 15 skipped
• python benchmarks/bench_kda_sm90_prefill.py --mode both
• python benchmarks/bench_kda_sm90_cp.py

[cherhh]

LGTM.