Support runtime stats for tiflash

[guo-shaoge]

Enhancement

Runtime stats refer to recording the characteristics of a query during execution, and using these statistics to select a more optimal execution strategy the next time the same query runs. Typical runtime stats include:

1. Record the average length of the key for HashAgg on key_string. Based on this length, an appropriate HashTable can be chosen. For example, if the average length is greater than 24B, a regular HashTable is used; otherwise, a StringHashTable is selected (by using norma HashTable, we can support prefetch. because StringHashMap doesn't support prefetch) (major, 10%).
2. Whether the query last used a two-level HashMap for HashAgg. If so, the two-level HashMap can be used directly in the next execution, saving the time of converting from one-level to two-level. (minor, <=5%)
3. Based on the size of the HashTable in HashAgg, an early resize may be performed (though this is harder to implement in TiFlash, as the region selected for a query may change due to load balancing strategies).
4. The reserve size for ColumnString. By default, it is assumed that the string length is 64B. If more accurate information becomes available later, it will improve the expression evaluation. (minor, <=5%)
5. When the NDV (Number of Distinct Values) in HashAgg is very low, using mini-batch may have side effects. Through runtime stats, the non-mini-batch version can be proactively chosen for the next query execution. (minor, <= 5%)