Improve current slow-store detecting

## Development Task

The current slow-score solution has some limitations:
* Huge applying cost: need to config the timeout-threshold value manually for each cluster
* Can't react by cluster status: it might do nothing if the threshold is not reached even when QPS rapidly drop
* Hard to config: too small will cause fake-possitive. Can’t detect anything if it’s too large (default value is too large)
* Need to manually tune the threshold value to fit vary hardwares and workloads for every cluster
* Can’t apply when some nodes are always slower than the others (eg, caused by workload tilting, which is very common)

To address issues above, we need a better slow-store detecting.