Change cwrsi() to operate on rows of U instead of columns.

It is no slower with a large number of pulses, and as much as 30% faster with
 a large number of dimensions.