Add DeepSeek-V4 H20 prefill kernel benchmark data

[xuqianmamba]

Summary

• Add H20 DeepSeek-V4 sparse-prefill attention measurements for attention
  TP2/DP4 and TP8/DP1.
• Add the H20 TP8 DeepSeek-V4 MoE gate/up and down measurements to the
  existing grouped-GEMM data table.
• Preserve every field and every row from the collected H20 kernel profile
  in topology-specific key_kernels.csv files.
• Add a small benchmark that calls the same sparse FlashMLA API used by
  SGLang instead of reimplementing the kernel.

Data layout

The attention files follow InferSim's existing attention CSV schema exactly:

dtype,s_q,s_kv,latency_us,mfu

The topology and kernel variant are encoded in the path:

• tp2-dp4/attn-64-64-c4.csv
• tp2-dp4/attn-64-64-c128.csv
• tp8-dp1/attn-16-64-c4.csv
• tp8-dp1/attn-16-64-c128.csv

The two MoE kernel measurements use InferSim's existing grouped-GEMM schema:
up_proj is the fused gate/up kernel and down_proj is the down kernel.
The measured runtime used flashinfer_mxfp4 (W4A16).

The complete profile data is also retained without dropping any metrics:

• attention_tp2_dp4_moe_tp8/key_kernels.csv: four TP2/DP4 attention rows
  covering C4/C128 at 4K and 8K.
• attention_tp8_dp1_moe_tp8/key_kernels.csv: two TP8 attention rows and
  the MoE gate/up and down rows at 4K.

Both files retain all 12 source columns: phase, component, workload,
topology, calls, average latency, FLOPs per call, achieved TFLOP/s, MFU,
traffic estimate, effective bandwidth and bandwidth utilization.

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

Attention FLOPs and MFU

Sparse-attention FLOPs are counted as:

FLOPs = 4 * s_q * local_heads * 64 * effective_keys

The factor 4 includes QK and PV (2 FLOPs per multiply-accumulate for each).
For attention TP2, local_heads = 128 / 2 = 64; for TP8,
local_heads = 128 / 8 = 16. The other 64 is the actual DSV4 sparse
attention compute width per head, not the wider 512-dimensional projection
workspace.

For example, TP2/DP4 C4 at 4K uses:

4 * 4096 * 64 * 64 * 5120 = 343.597 GFLOPs
343.597 GFLOPs / 2914.016 us = 117.912 TFLOP/s
MFU = 117.912 / 138.184 = 0.8533

MFU is stored as a fraction, matching the existing InferSim CSV files. The
denominator is the measured H20 BF16 GEMM peak of 138.184 TFLOP/s.

Validation

• python3 -m py_compile kernel_benchmark/flashmla_dsv4_sparse_prefill.py
• benchmark --help smoke test
• exact CSV-header validation against
  bench_data/dsa/prefill/h800/attn-128-576-2048.csv
• exact row-by-row comparison confirming that all 12 columns and all eight
  rows from the collected H20 kernel data are present
• 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.}

[xuqianmamba]

Superseded by #17. The replacement PR uses the kernel-oriented bench_data/dsa layout and encodes TP/DP through local head counts instead of topology subdirectories.