support mpp err recovery

[guo-shaoge]

Enhancement

In disaggregted tiflash situation, AutoScaler assigns init number of CN to one tenant(normally 1) to handle MPP query.
But when user run a big query which one node cannot handle(e.g. MemLimit exceeds), the query will give error.

But actually we make AutoScaler scale out immedicately and retry this query internally, so the query will run successfully.