[NSE-17]TPCDS Q72 optimization

[xuechendi]

couple optimization has been made:

1. if hash map key is unique, we will directly put its ArrayItemIndex in key map instead of link table,
2. hashmap size will be predicted by its total size, so small input will only allocate small size hash table
3. move hash calculation inside Get/IfExists function and add a cache their to fast fetch most recent probed key.
4. switch smj left and right sequence upon its numOutputFields.
5. skip is_null check in payload column
6. fix a small issue in round function

verified with TPCDS_DPP_SF500_V1

Fixed: #17

[github-actions]

[xuechendi]

@zhouyuan， verified with TPCH sf500 as well, I think we can merge it.