Symmetrical matrices: compute each bench only once in `compute_benchs`
In benchmarks.compute_benchs
, the benchmark matrices to be computed are symmetrical, but we are still doing both benchmarks. With the current caching strategy, it is inconvenient to let the cache handle this redundancy — processes cannot that easily back-propagate and unify the cache in the main process.
The best way to handle this would simply be to skip the lower part of the matrix, when symmetrical.