F3 — CUDA backend

[kkokosa]

A GPU backend for the same .dfb checkpoints and the same semantics as the CPU backends.

Deliverables

• PTX kernels for phase A (LIF update, threshold, refractory countdown) and the blocked-CSR delivery; driver-API binding (no CUDA toolkit dependency at run time beyond the driver).
• CUDA graphs for the 18-step batch; lanes (batched trials) on the GPU.
• Cross-backend validation: spike-identical to the CPU reference on the golden fixtures and statistically equivalent (per-neuron spike counts) on the full v630 sugar run; determinism for a fixed seed.
• dotfly bench --backend cuda numbers for the three standard scenarios.

Acceptance: ≥ 10× real time for the full MaleCNS on one consumer GPU; Backend.Cuda selectable from SimulationOptions; the room demo runs unchanged on it.

Context: PLAN.md §11 (F3), §8 (performance design).