add kernel-arena reference solutions + restore FlashInfer GDN baseline (H100 PCIe)

Summary

Two contributions to the FlashInfer-Trace dataset, both verified on NVIDIA H100 PCIe.

1. Restore FlashInfer GDN baseline (was missing from traces/baseline/)

The traces/baseline/gdn/ directory did not exist in the dataset prior to this
PR. The FlashInfer GDN solution flashinfer_wrapper_9b7f1e (CuTe-DSL pretranspose
path) silently fails to import out-of-the-box on NGC nvcr.io/nvidia/pytorch:24.10-py3:

• cuda-python installs to a different location than NGC's system cuda regular
  package, so cuda.bindings is not visible after a normal pip install --target.
• nvidia-cutlass-dsl ships its package via a .pth file that pip install --target
  does not honor.

The result: flashinfer.gdn_kernels.run_pretranspose_decode is silently None
and all baseline runs return RUNTIME_ERROR.

After installing nvidia-cutlass-dsl[cu13] + cuda-python and applying a small
sitecustomize.py patch (see Reproducing the GDN baseline below), the
baseline imports and runs end-to-end. This PR adds 5 PASSED baseline traces
at 33.03×–35.16× speedup vs PyTorch reference for
gdn_decode_qk4_v8_d128_k_last.

2. Add kernel-arena reference solutions (SDPA, FLA)

Two third-party-library reference solutions, useful as comparison anchors:

Honest framing: these reference solutions are slower than the FlashInfer
baselines. On GDN, FLA reaches ~12× vs PyTorch ref versus FlashInfer's ~33×
(FLA ≈ 0.36× of FlashInfer baseline). On GQA, SDPA is faster than FlashInfer
on short KV but ~50× slower on long KV. They are submitted as reference points
for community comparison, not as performance winners — FlashInfer's existing
hand-tuned kernels remain SOTA on these definitions.

Files added

• traces/baseline/gdn/gdn_decode_qk4_v8_d128_k_last.jsonl (5 PASSED, NEW directory)
• solutions/kernel-arena/gqa_paged/gqa_paged_decode_h32_kv8_d128_ps1/sdpa_paged_decode_v1.json
• solutions/kernel-arena/gdn/gdn_decode_qk4_v8_d128_k_last/fla_gdn_decode_v1.json
• traces/kernel-arena/gqa_paged/gqa_paged_decode_h32_kv8_d128_ps1.jsonl (48 PASSED)
• traces/kernel-arena/gdn/gdn_decode_qk4_v8_d128_k_last.jsonl (5 PASSED)

Suggested follow-up for the web viewer

For the kernel-definition page to render the now-available baseline, the
viewer's baseline map needs one extra entry. Maintainers may want to add to
flashinfer-ai/flashinfer-bench — web/apps/web/data/baselines.ts:

gdn_decode_qk4_v8_d128_k_last: {
  default: "flashinfer_wrapper_9b7f1e",
},

Without it, the GDN definition page shows "Baseline not available" even though
the traces in this PR are valid.

Environment

NVIDIA H100 PCIe · NGC nvcr.io/nvidia/pytorch:24.10-py3 · torch 2.11.0+cu130
· flashinfer-python 0.6.9 · flashinfer-bench (latest from PyPI) ·
flash-linear-attention 0.5.0 · nvidia-cutlass-dsl 4.4.2 [cu13] · cuda-python 13.2.0

Reproducing the GDN baseline

# sitecustomize.py — placed on PYTHONPATH; auto-imported by site.py
import sys, os
PIP_TARGET = '/path/to/pip-target'  # wherever you ran `pip install --target`
try:
    import cuda
    if os.path.exists(f'{PIP_TARGET}/cuda/bindings') \
            and f'{PIP_TARGET}/cuda' not in cuda.__path__:
        cuda.__path__.insert(0, f'{PIP_TARGET}/cuda')
except ImportError:
    pass
cutlass_path = f'{PIP_TARGET}/nvidia_cutlass_dsl/python_packages'
if cutlass_path not in sys.path:
    sys.path.insert(0, cutlass_path)

Then:

flashinfer-bench run --local . \
    --definitions gdn_decode_qk4_v8_d128_k_last \
    --solutions flashinfer_wrapper_9b7f1e fla_gdn_decode_v1 \
    --warmup-runs 5 --iterations 20 --num-trials 1