Floor residuals at tolerance

[stephane-caron]

Residual tables are normalized by the best solver's shifted geometric mean. When the best residual is very close to zero, that divisor yielded large or even negative values depending on the machine running the benchmark, as for instance in this report:

| ecos  | -218763017610834.0 | -225181.0 | -225173226328.0 | -225179973.0 |
| highs | -562949953421310.0 | -562951.0 | -562949953418.0 | -562949947.0 |

Since #136, qpbenchmark floors residuals at their corresponding per-settings tolerance before averaging. This way, solvers are not rewarded for solving more precisely than requested, and negative outputs are fixed.