Compute block-residuals, instead of scalar ones

This transverses the matrix less, and uses the cache better.  On an example
with 6x6 blocks I can measure a speedup in the range of 30% for a full
multi-grid iteration (including transfer operators, too).