disk-based partialResultMap for HashAggExec

[XuHuaiyu]

Feature Request

Is your feature request related to a problem? Please describe:

The main memory space consumption of HashAggExec is HashAggExec.partialResultMap which is of type aggPartialResultMapper.
https://github.com/pingcap/tidb/blob/master/executor/aggregate.go#L41

To avoid OOM in HashAggExec, we need a disk-based aggPartialResultMapper.

Describe the feature you'd like:

Implement a disk-based aggPartialResultMapper, partialResultMap should be spilled over to disk when mem-quota-query in tidb.toml is exceeded.

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy:

[SunRunAway]

Another approaching is like grace hash join.
Shuffle multiple parts of data, spill them into disk, then execute each part in memory.

[SunRunAway]

We should finish the memory tracking problem(#14103) before this issue

[SunRunAway]

Q18 cannot finishes w/ TPCH 10g workload with a 16GB memory instance without this feature.