Optimize HashAgg by caching collation sort keys for repeated string keys

[ChangRui-Ryan]

Enhancement

For string GROUP BY keys using a non-trivial collation such as
utf8mb4_general_ci, HashAgg currently generates a collation sort key for
every input row.

When the input has many repeated raw string values, the same sort key is
generated repeatedly. For example:

SELECT l_shipmode, COUNT(*)
FROM lineitem
GROUP BY l_shipmode;

If l_shipmode uses utf8mb4_general_ci and has a low NDV, sort-key
generation can consume a significant portion of the aggregation time.