Add support for projection pushdown into struct fields

[Mytherin]

This PR adds support for projection pushdown into struct fields, allowing for scans to only partially scan structs - the parts that are required for the query. The way this is supported is that there is a new class - ColumnIndex - that supports recursively qualifying which columns are needed for the query:

struct ColumnIndex {
	idx_t index;
	vector<ColumnIndex> child_indexes;
};

For example, if we are querying a struct STRUCT(a INT, b INT), and we only need to access the second struct element (b) - the child index would have index = 1.

The child ColumnIndexes are set by the RemoveUnusedColumns optimizer - which looks for calls to struct_extract or array_extract and figures out which child indexes are required to complete the query.

Backwards Compatibility

The vectors that are returned by the scans are still in the same form - i.e. the scan still returns a STRUCT(a INT, b INT). The only difference is that we don't scan the sub-columns we don't need - i.e. the column a INT is filled with a constant NULL value. In that way, scanning the sub-columns we don't need is actually optional. This allows for the struct projection pushdown to be fully backwards compatible.

The table function receives the column indexes it needs as input in the TableFunctionInitInput. The old column_ids are still present as well, which can still be used by existing table functions.

struct TableFunctionInitInput {
	vector<column_t> column_ids;
	vector<ColumnIndex> column_indexes;
};

Benchmarks

Below are some benchmarks, running various TPC-H queries over SF10. For the struct case, we store the entire rows in a struct field, and then extract it using a view, e.g.:

CALL dbgen(sf=10, suffix='_normalized');
CREATE TABLE lineitem_struct AS SELECT lineitem_normalized AS row_val FROM lineitem_normalized;
CREATE VIEW lineitem AS SELECT UNNEST(row_val) FROM lineitem_struct;

Query	Struct (v1.1)	Struct (new)	Regular
Q01	0.33s	0.15s	0.14s
Q06	0.26s	0.07s	0.04s
Q09	1.17s	0.62s	0.2s

As we can see, with the projection pushdown the performance is mostly the same between a regular table scan and scanning unnested struct columns. There are still some differences in the generated query plans (different join order in Q09) - most likely caused by either some missing statistics or statistics not being propagated correctly, to be perhaps investigated in the future.

Future Work

This PR only implements struct projection pushdown for DuckDB's native table scans - it still needs to be implemented for the parquet reader.