tiflash scale in is slow

[guo-shaoge]

Enhancement

For now, when AutoScaler want to scale in one CN, supervisor will send SIGKILL to tiflash process. So tiflash has no chance to do any cleanup work, most importantly, the LAC::stop(), which will send final report rpc to PD to cleanup some slot info.

We can send SIGTERM to let tiflash do its cleanup, but for now tiflash takes almost 10sec from it receive SIGTERM to exit properly, which is too long for AutoScaling. (check the log rec_stop.log.md)