[NSE-857] Fill destination buffer by reducer

[FelixYBW]

What changes were proposed in this pull request?

The solution is to create an offset array which list the src offset for each reducer, like below:
reducer0:

reducer#	Vector index	Vector value
0	0	0
0	1	4
0	2	6
1	3	1
1	4	5
1	5	8

Then we can read the src column randomly and fill the destination column sequentially one reducer by one reducer. The source data size should be smaller enough to hold into L1/L2 cache, and make sure source and destination cache line both are read onece. Otherwise the performance will be very bad. Currently the recordbatch size is 32K rows, so for double column the size is 128K.
On the write, we can use NTStore to bypass RFO. Then we can avoid the cache polution. But when reducer# is very large, ntstore doesn't works well because each reducer will be only fill little data, like 32K batch for 4000 reducer, each reducer will be written 8 values only.

How was this patch tested?

From benchmark data, the solution partially solved the reducer# scaling issue. From below chart, we can see 4096 and 512 reducer has the same performance

Remining work

AVX implementation doesn't show better performance than INT solution. NTStore either doesn't show better performance

[github-actions]

#857

[FelixYBW]

Jenkins TPCH done. 8x partition doesn't show any regression now.
tpch_2022_05_01_application_1647347981137_0322.html

Note it doesn't solve the reducer# scaling issue completely. With even larger reducer#, each reducer only gets a few rows which doesn't fill up 64Byte cache line. So the memory throughput vs. split data size ratio will increase again and performance will drop soon. The another solution is to convert into row based format during split. At recucer side, return to columnar format.

[FelixYBW]

To record, Original performance:
tpch_2021_12_01_application_1638156867030_0038.html