Predicate pushdown using Parquet row-group statistics

Real engines don't read data they can prove is irrelevant. Parquet stores per-column min/max statistics for each row group, so a selective predicate can skip whole row groups before reading them. This is what Spark's PushedFilters and row-group pruning do.

Task

When a Scan carries a pushdown predicate (new_scan(path, pushdown=Compare(Col("id"), ">", Lit(950)))), skip any row group whose statistics prove no row can match, and record how many you skipped.

• Read stats via pf.metadata.row_group(rg).column(idx).statistics (has_min_max, .min, .max) — no data read required.
• Maintain the counters on ScanStats: row_groups_total, row_groups_read, row_groups_skipped.
• Surface them in explain(): a PushedFilters=... annotation and the read/skipped counts.

Soundness (the one rule you must not break)

Skip a row group only when the predicate is impossible on [min, max]:

If statistics are missing or the predicate isn't a simple Col op Lit, do not skip (conservative). Pushdown only prunes row groups — keep the Filter operator so surviving rows are still filtered exactly.

Checks

• A selective predicate over a sorted file skips most row groups (counter).
• explain() shows PushedFilters and skipped counts.
• Correctness is unchanged vs DuckDB for every threshold (Hypothesis property): a skipped group must never have contained a matching row.