feat(benchmark): Add unified benchmark suite

[m96-chan]

Summary

• Add modular benchmark suite under src/pygpukit/benchmark/
• Unified API for GEMM, GEMV, and attention benchmarks
• JSON export and baseline comparison with regression detection
• CLI interface: python -m pygpukit.benchmark

Features

BenchmarkSuite API

from pygpukit.benchmark import BenchmarkSuite

suite = BenchmarkSuite()
suite.add_gemm(sizes=[(4096, 4096, 4096)], dtypes=["bf16", "tf32"])
suite.add_gemv(dtypes=["bf16", "fp8", "nvf4"])
suite.add_attention(seq_lens=[512, 1024, 2048])

report = suite.run()
report.save("baseline.json")

# Compare with baseline
comparison = suite.compare("baseline.json", threshold=0.05)
if comparison.has_regression():
    raise RuntimeError("Regression detected!")

CLI Usage

python -m pygpukit.benchmark --quick
python -m pygpukit.benchmark --gemm --sizes 4096,8192
python -m pygpukit.benchmark --save results.json
python -m pygpukit.benchmark --compare baseline.json --fail-on-regression

Modules

File	Description
__init__.py	BenchmarkSuite class, exports
base.py	Base Benchmark class, measure_kernel()
results.py	BenchmarkResult, BenchmarkReport, ComparisonResult
gemm.py	GEMMBenchmark, FP8GEMMBenchmark
gemv.py	GEMVBenchmark, W8A8GEMVBenchmark
attention.py	SDPABenchmark, GQABenchmark
cli.py	CLI argument parsing and entry point

Test plan

• Build package and verify import
• Run python -m pygpukit.benchmark --quick
• Run python -m pygpukit.benchmark --save test.json
• Run python -m pygpukit.benchmark --compare test.json

Closes #163

🤖 Generated with Claude Code