Add DeepSeek-V4 H20 kernel benchmark data

[xuqianmamba]

Summary

• Add NVIDIA H20 sparse-prefill attention measurements under the existing kernel-oriented bench_data/dsa/prefill/h20 layout.
• Add the DeepSeek-V4 H20 MoE gate/up and down measurements to the existing bench_data/grouped_gemm/prefill/h20/data.csv table.
• Add a small benchmark that invokes the same sparse FlashMLA APIs used by SGLang instead of reimplementing the attention kernel.

Data layout

The attention CSVs use InferSim's existing schema:

dtype,s_q,s_kv,latency_us,mfu

TP/DP topology is not encoded in directories. The file names record the kernel shape:

• attn-16-64-c4.csv and attn-16-64-c128.csv: 16 local attention heads (attention TP8)
• attn-64-64-c4.csv and attn-64-64-c128.csv: 64 local attention heads (attention TP2)

The second 64 is the sparse-attention compute width per head, while c4/c128 identifies the compression ratio.

All measurements in this PR were collected on NVIDIA H20 GPUs.

Validation

• python -m py_compile kernel_benchmark/flashmla_dsv4_sparse_prefill.py
• benchmark --help smoke test
• CSV header validation against bench_data/dsa/prefill/h800/attn-128-576-2048.csv
• CSV parsing and non-empty row validation
• git diff --check

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}