scheduler: better expectation and variance in hot region scheduler

[lhy1024]

Development Task

In 4.0, we use expectation to avoid creating redundant operators after #2297

In 6.3, we use variance to reduce schedule when the cluster has been enough uniform after #4912 and #5286

But we calculate expectation and variance without considering about placement-rules or machine performance, only distinguishing tikv and tiflash.

we need better expectation and variance in hot region scheduler, or could close then they can not work.

• add config
• uniform toleranceRatio, stddevThreshold and balancedRatio
• consider about label