Move up the specialization of std::numeric_limits

This fixes a compilation error seen when building TensorFlow on macOS:
https://github.com/tensorflow/tensorflow/issues/17067
1 file changed