Storage: Lazily Load Row Groups from Tables

[Mytherin]

This PR modifies the initial startup of a database so that row groups are lazily constructed, instead of eagerly constructed. This is generally the largest contributor to start-up cost of large databases as the amount of row groups grows along with the size of tables. The lazy row group construction is done in the SegmentTree structure. Whenever a row group is requested from the segment tree, row groups are loaded as needed. This also means that not necessarily all the row groups have to be loaded. For example, when running a LIMIT on a table, we only need to load the initial row groups.

Below are some timings run on a database file containing a single large lineitem table at scale factor 120 (or rather, 120x SF1 lineitem). This is around 770 million rows, or 6K row groups.

Query	v0.6.1	v0.7.1	master	new	Parquet
SELECT 42	1.60s	0.31s	0.21s	0.02s	-
FROM lineitem LIMIT 1;	1.62s	0.32s	0.22s	0.03s	0.27s

Segment Tree Restructuring

This PR restructures the segment tree so that it is fully templated, and also adds templating to the SegmentBase. This avoids the need for casting to and from the SegmentBase class everywhere. RowGroup now inherits from SegmentBase<RowGroup> and there is a RowGroupSegmentTree that inherits from SegmentTree<RowGroup>.

Rather than directly accessing the next pointer of the SegmentBase, the SegmentTree:: GetNextSegment is now always called. This will usually still follow the next pointer (as that is the most efficient) - but might trigger the lazy loading of segments if necessary.

In addition, other methods (like GetRootSegment, GetSegmentByIndex, etc) are also overloaded to correctly trigger lazy loading of row groups if required.

Zone Maps Optimization

This PR also adds optimizations around the pushdown of zonemaps to avoid value construction in the NumericStats. This speeds up queries whose query time is dominated by zonemap skipping significantly, which can happen if we are selecting a small subset of data from a large table with a lot of row groups.