pub fn compile_to_cuda_tensor_op(source: &str) -> Result<String, String>Expand description
Recognize ANY supported tensor op — matmul OR elementwise (plus/hadamard) — in a LOGOS program and
lower it to CUDA. The unified accelerator entrypoint: tries the GEMM idiom first, then the elementwise
idiom, and REFUSES (typed) only if neither is present. This is what closes the tensor-op × backend matrix
to zero gaps on the CUDA cell.