Carpet already assumes that these grid scalars have the same value on all processors. So reducing them does not generate any information that is not already present by just looking at a single value.
The attached patch outputs grid scalars (from the root processor) directly, this means that grid scalar output files can be used to output timeseries like grid scalars (which is what many are).
The attached parfile and data files are a modfied version of a QLM test. Notice how much nicer to read the scalar output .scalar.asc is then the 0D output ..asc.