Tensor: Enable double outer reductions in the measured shape range

Match the float/double capability flag with a constrained launch overload and typed output initialization. Use the generic evaluator for double outside reductions of at least 64 elements with 8 to 64 outputs per multiprocessor.

Add correctness cases at the dispatch boundaries and a reduction benchmark against CUB. Validate outputs against compensated host references with statistical tolerances for the benchmark's random inputs. Request C++17 for CUDA 13's CUB while retaining Eigen's C++14 header contract.