Function active_backend 

pub fn active_backend() -> Backend

The backend dense matmuls will actually use in this process, decided by the same cached env/probe reads dispatch uses. CUDA wins when both are compiled in, matching [WeightMatrix::apply_gpu]’s order.