GPU Napkin Math v0.1.0

Initial release of GPU Napkin Math: measured GPU numbers and a small roofline-style estimator inspired by sirupsen/napkin-math.

Highlights:

• Reproducible Modal benchmark pinned to an H100 SXM
• CUDA-event timing with warmups and seven retained samples
• Correctness validation for transfers, memory operations, and GEMMs
• Rounded H100 napkin numbers for launch cost, PCIe, HBM, FP16, BF16, TF32, and FP32
• Installable gpu-napkin workload estimator

The reference repository is not included.