First real-GPU results

This file was the placeholder for "acceptance numbers once a GPU box exists." That happened: the offload path was validated end-to-end on an RTX 2060 (sm_75) — device probe, transparent --gpu offload, checksum parity with HotSpot, and head-to-head numbers against TornadoVM 4.0.1 (PTX).

The living results now belong to:

• bench-gpu/results/ — dated benchmark tables (cold 4-way comparison, warm-loop comparison, div-chain "GPU vs best CPU" concept-prover).
• Repo README.md → "GPU offload" — the headline table.
• docs/gpu/reductions.md — intentional CPU fallbacks (float reductions) and their rationale.

Acceptance criteria from the original plan, as measured (warm, full H2D + kernel + D2H per call, GpuWarm.heavy n = 2²⁴): correctness ✓ (samples/checksums match HotSpot), speedup ✓ (GPU ≥ 2× target — measured ~475× vs CratonVM CPU JIT and ~1.1–2× vs TornadoVM on the same kernel).