Fixes T7664. When there are a large number of tasks (400+) with the same subpriority (which can happen if the subpriority features are rarely used), it may take more than 30 seconds to rebalance them.
Make the algorithm more aggressive about rebalancing homogenous blocks of tasks.
This may need to get even fancier, but I'd guess it can process blocks 1-2 orders of magnitude larger, which should be ~all installs.
(If someone still hits issues with this, I'll make it fancier.)
Once a block is rebalanced, it doesn't need to be rebalanced again (at least, not as a whole block) so we basically just need to get over the initial hurdle here and then we're good.
In the worst case, we can provide bin/maniphest rebalance or similar and do the rebalance step offline.
And, in any case, we have more test coverage here now.