Tune the perfomance of hash aggregate operator

[XuHuaiyu]

After #7268 , all the aggregate evaluation was switched to new the evaluation
framework, but there still remains some optimization to be done to speed up
the evaluation, reduce the memory usage or cache miss ratio. We need to test
the new framework through some benchmarks, and profile the execution process
to find the optimize points.

related issues:
#6952

The details of performance tuning are shown in the following comments, and the
optimization points we could focus on are listed as check list in every comment.

Note: These optimizations are planned to be finished before 2018/09/09 CST,
it would be nice if you can push your pull requests and get them merged before
that deadline. ^_^

[XuHuaiyu]

T1:

DataSet	row count	group count
TPC-H 50G	75,000,000	5,000,000

sql	flame graph	profile graph
select avg(o_totalprice) from orders group by o_custkey;	q1.txt	q1-torch.txt

Note: the name suffix of flame graph and profile graph should be changed to .svg

explain:

mysql> desc select avg(o_totalprice) from orders group by o_custkey;
+-----------------------+-------------+------+-------------------------------------------------------------------------+
| id                    | count       | task | operator info                                                           |
+-----------------------+-------------+------+-------------------------------------------------------------------------+
| HashAgg_9             | 4973568.00  | root | group by:col_2, funcs:avg(col_0, col_1)                                 |
| └─TableReader_10      | 4973568.00  | root | data:HashAgg_5                                                          |
|   └─HashAgg_5         | 4973568.00  | cop  | group by:tpch50.orders.o_custkey, funcs:avg(tpch50.orders.o_totalprice) |
|     └─TableScan_8     | 75000000.00 | cop  | table:orders, range:[-inf,+inf], keep order:false                       |
+-----------------------+-------------+------+-------------------------------------------------------------------------+
4 rows in set (0.00 sec)

Optimize point already found:

• Define HashAggFinalWorker.groupSet as stringSet instead of MVMap
• Define HashAggIntermData.groupKeys as []string

Through the above optimization, we can avoid the cost of:

1. 3.77% runtime.stringToByteSliceof CPU cost introduced by HashAggPartialWorker.shuffleIntermData
2. 0.75% runtime.sliceToString of CPU cost introduced by baseHashAggWorker.getPartialResult
3. 8.86% MVMap.Get and MVMap.Put of CPU cost introduced by access to MVMap(we may introduce new cost of access to stringSet which would cost less than before)

[XuHuaiyu]

T2:

DataSet	row count	group count
TPC-H 50G	75,000,000	5,000,000

sql	flame graph	profile graph
select avg(distinct o_totalprice) from orders group by o_custkey;	q2.txt	q2-torch.txt

Note: the name suffix of flame graph and profile graph should be changed to .svg

explain:

mysql> desc select avg(distinct o_totalprice) from orders group by o_custkey;
+---------------------+-------------+------+----------------------------------------------------------------------------------+
| id                  | count       | task | operator info                                                                    |
+---------------------+-------------+------+----------------------------------------------------------------------------------+
| HashAgg_5           | 4973568.00  | root | group by:tpch50.orders.o_custkey, funcs:avg(distinct tpch50.orders.o_totalprice) |
| └─TableReader_9     | 75000000.00 | root | data:TableScan_8                                                                 |
|   └─TableScan_8     | 75000000.00 | cop  | table:orders, range:[-inf,+inf], keep order:false                                |
+---------------------+-------------+------+----------------------------------------------------------------------------------+
3 rows in set (0.00 sec)

Optimize point already found:

• Define a Datum slice buffer in HashAggExec to replace the definition of vals in HashAggExec.getGroupKey

Through this optimization, we can avoid the cost of:
4.5% runtime.makeslice introduced by the definition of vals in HashAggExec.getGropuKey

• Define HashAggExec.groupMap as stringSet instead of MVMap

Through this optimization, we can avoid the cost of:
17.5% MVMap.Get and MVMap.Put introduced by the access to MVMap, but it'll bring new cost of runtime.byteSliceToString and access to stringSet, so we need to verify whether this optimization is beneficial.