Performance read of short and wide sparse matrices

Since for sparse matrix reads of many columns we synchronize over the first row of a row of blocks, there is unnecessary contention with many read threads and a small number of rows (e.g., 10K rows -> 10 blocks, but 24 threads reading)