Use a set instead of vec to represent the group-by clause in the Agg operator

[chenzl25]

          > > Theoretically, the ordering in group-by clause doesn't matter, however, in our LogicalAgg implementation, we store them in a vec so the ordering matter in terms of common sub-plan detection.

Does this mean that extracting common plan node with Eq is too strict? It'll be nice if we can share Agg in this case with a Project automatically inserted.

Actually, I prefer to replace the group-by vec with abitmap.

Originally posted by @chenzl25 in #8159 (comment)

The advantages of using a set:

• Dedup group key automatically.
• Provide more opportunities for common sub-plan sharing optimization.

[liurenjie1024]

Or an ordered vec?

[chenzl25]

Or an ordered vec?

The ordered vec might have duplicate keys.

[liurenjie1024]

LGTM

[st1page]

Sometimes group keys matter e.g. we need to permute the group keys with watermarks #6452. or we need to break some assumption of the streaming executor( the group key is not the fitst N columns in the state table's primary key).

[chenzl25]

Sometimes group keys matter e.g. we need to permute the group keys with watermarks #6452. or we need to break some assumption of the streaming executor( the group key is not the fitst N columns in the state table's primary key).

Yes, the group-by set defines the schema of the Agg operator, but we can choose to implement the state table with a primary key in any order you want e.g. based on the watermark information.