Add fused SwiGLU kernel benchmark

[WaffleBits]

What changed

• add a fused Triton SwiGLU kernel with FP32 PyTorch oracle validation
• autotune block size and warp count per tensor size
• benchmark RMSNorm and SwiGLU through the same cache-controlled CUDA-event harness
• commit a 12-case RTX 5070 Ti report with raw samples and regression-compatible case keys
• document measured results and the boundary between logical bandwidth and hardware counters

Why

Current inference performance roles consistently ask for custom kernels, correctness-first optimization, low-precision numerics, launch tuning, and quantitative evidence. This broadens the lab beyond a single reduction kernel without inventing scale claims.

Validation

• python -m pytest
• python -m ruff check .
• python -m compileall -q src tests
• full RTX 5070 Ti cache-cold run: 100 warmups, 500 samples, FP16/BF16, three shapes, both kernels
• all 12 GPU correctness cases passed