v3: fix pre-Hopper (32MiB, L2] launch cliff (independently reported); retract position-biased 4090 claims; buffer-fair dtype benchmarks (MK, powered by Claude)

v2: H100 (sm_90) retune — Hopper sub-16MB launch branch, H100 benchmarks/docs, version bump (MK, powered by Claude)

Optimized relu: cpu/cuda/xpu, 1.2-1.85x faster on RTX 4090, benchmarked vs upstream and torch.relu