Epic: duckdb performance parity with Parquet

[myrrc]

In other words, where and why are we not better than Parquet on all benchmarks we check:

clickbench: q23[1], q35, q40, q41, q42.
statpopgen: q2, q3, q4, q5, q9, q10.
tpch S3 sf=1: q6, q17, q21.
tpch NVME sf=10: q4, q12 (vortex-compact).
tpch NVME sf=100: q20, q21.
tpch S3 sf=100: q6, q17, q21.
tpcds NVME sf=1: many, notable ones are q2, q3 (big regression), q5, q6

Features that will solve this:

• [1] Lazy materialization: Late materialization support for duckdb #7721
• [2] Post-filter cardinality estimates: remove pushed filters so Duckdb doesn't apply them again: ji/remove-double-filter #7315
• [3] Partition stats Add file-wise partition stats for duckdb #7615
• [4] Expression pushdown Expression pushdown for duckdb #7727