duckdb_pglake: parallelize postgres_scan on partitioned tables

[sfc-gh-mslot]

Problem

When scanning a partitioned PostgreSQL table, postgres_scan previously fell back to a single serial task with one thread and zero parallelism. Because partitioned parent tables have no physical heap storage (relkind = 'p'), their relation size is 0 pages. Furthermore, parallelizing a partitioned table directly by splitting the parent's ctid range is invalid: each physical partition has its own independent block numbers starting at block 0, so a ctid range filter on the parent evaluates against every partition simultaneously, concentrating rows in the first task and returning empty results for later tasks.

Solution

Add table-partition-scan.patch to duckdb-postgres to make postgres_scan and attached partitioned table scans partition-aware:

1. Hierarchy discovery: Recursively query pg_inherits to resolve all underlying leaf partitions (relkind IN ('r', 'm', 'f')), handling arbitrary subpartitioning depth as well as partitions residing in separate schemas.
2. Cardinality and thread estimation: Set the table's pages_approx to the sum of all leaf partition page counts (pg_relation_size(oid)), yielding accurate cardinality estimates and sizing max_threads across all partition tasks.
3. Partition-level and ctid-level parallelism: Parallel tasks are scheduled across partitions. Partitions larger than pg_pages_per_task are subdivided into ctid-range chunks within the partition, while smaller or empty partitions are scanned as single tasks.
4. Chunk completion fix: Fix ScanChunk to return non-empty output buffers when a task finishes with remaining tuples, preventing the next task from overwriting unconsumed results.

-- Scan a partitioned table in parallel across and within partitions
SELECT count(*) FROM postgres_scan('host=localhost dbname=mydb', 'public', 'measurements');

-- Pushdown filters and projection to underlying partitions
SELECT id, value FROM postgres_scan_pushdown('host=localhost dbname=mydb', 'public', 'measurements') WHERE id > 1000;

Test plan

• Tested range-partitioned table scans returning all rows from leaf partitions.
• Tested multi-level subpartitioning hierarchies with partitions in separate schemas.
• Tested empty partitioned tables (no partitions attached) returning zero rows without error.
• Tested parallel ctid chunking across and within partitions with pg_pages_per_task = 5.
• Tested filter pushdown and projection pushdown on partitioned tables.
• Tested binary and text protocols across multiple tasks without lost tuples.
• Verified full test suite in pgduck_server/tests/pytests/test_postgres_scanner.py (30 passed).

[sfc-gh-dachristensen]

Is there a similar issue with table inheritance?

[sfc-gh-dachristensen]

approx_num_pages is a proxy for being a branch partition?