Add lazy physical parquet scans

[platypii]

Summary

Adds parquetScan, a lazy, column-oriented API for query engines that need physical row ranges and selective column reads.

What changed

• Exposes candidate row ranges using absolute, half-open file coordinates.
• Reads individual columns lazily for a candidate range or exact subrange.
• Uses row-group statistics, bloom filters, and page indexes for conservative pruning.
• Keeps pruning separate from exact value filtering.
• Enables offset-index range reads by default for scans.
• Automatically makes pruning-filter columns available without eagerly reading them.
• Caches covering column reads for contained subranges.
• Shares metadata, index preparation, and physical-plan execution with existing read APIs.
• Adds public TypeScript types and exports for the scan API.

Performance

Scan preparation reads metadata and requested indexes only; column data remains lazy. Exact subranges use offset indexes where available to avoid reading complete column chunks. Existing row-oriented reads retain eager, parallel range fetching.

Validation

• Added coverage for lazy reads, exact subranges, cache reuse, pruning, absolute coordinates, and API validation.
• Verified offset-index reads transfer less data than full-column reads.
• Verified filter-only columns are not unnecessarily fetched by parquetReadColumn.
• Tests, lint, generated types, fixture comparisons, and differential checks against master pass.

Follow-up opportunity

The scan currently caches the most recent requested range per column. Retaining page-aligned decoded coverage and parsed offset indexes could avoid repeated I/O for adjacent, separately requested subranges.