ferrox-cuda

CUDA/NVRTC kernels for Ferrox.

Capability detection always builds. Real CUDA dispatch sits behind --features cuda. Loading is dynamic, through cudarc, so no toolkit is needed to compile. Run the hardware tests only on a machine with a GPU.

mul_mm, the batched quantized GEMM, is unrun on hardware and not wired into any model path yet. It ships with a scalar twin (mul_mm_ref) that the default cargo test -p ferrox-cuda exercises, and with tools/mul_mm_host_check/run.sh, which compiles the emitted CUDA C against a barrier shim and runs it on this CPU to check it against that twin. Neither is a measurement. Nothing in docs/ may claim a CUDA GEMM capability until cargo test -p ferrox-cuda --features cuda -- --ignored has run on a real device.