Function compile_to_cuda_matmul_with 

pub fn compile_to_cuda_matmul_with(
    source: &str,
    knobs: &CudaKnobs,
) -> Result<(MatmulSpec, String), String>

Knob-driven counterpart of compile_to_cuda_matmul — recognize the matmul, then lower it under an explicit crate::codegen_cuda::knobs::CudaKnobs point in the optimization lattice (e.g. the Int8Dp4a precision fast path). This is the user-facing surface the CLI/API drive to select a trade-off; a non-matmul is refused (typed) regardless of the knobs.