datalake_fdw: read path — planned fragments, CustomScan, projection and pruning

Start with the engine plan and the CustomScan planner hook, using dependencies A, B2, and B4; validate the skeleton, projection, and pruning against the stub and local files first. Done means matching Spark rows across schema evolution and fragment splits, reporting correct EXPLAIN ANALYZE pruning, refusing unsupported or dictionary-encoded columns by name, and releasing readers on cancellation.

Part of #2008. Letters (A, B0–B7, C, D, E) are the PRs listed there; this is D.

Scope

• Fragments from the engine's plan (files, row-group ranges) assigned to segments; one large file may be split by row group.
• A CustomScan (planner hook, design decision D10) that opens a reader per fragment and decodes batches into slots.
• Projection by field id from the table's Iceberg schema (ProjectionSet), so files written before a column was added or renamed read correctly; only needed columns are read.
• Row-group pruning from the query's quals against Parquet statistics; EXPLAIN ANALYZE reports row groups skipped.
• With it: #1989 name mapping for files without field ids.

Out of scope

Delete files (E), time travel (#1683 §2.3), ANALYZE beyond the current zero-sample no-op.

Depends on

A, B2, B4. Skeleton, projection and pruning are testable against the stub and local files first.

Acceptance

• Same rows as Spark on the same table, including one Spark evolved (column added, renamed, dropped, int promoted to long).
• A file split into three fragments reads the same as whole; every segment reads only its fragments.
• EXPLAIN ANALYZE shows pruning for a selective predicate and none for WHERE true.
• Unsupported or dictionary-encoded columns are refused naming the column; cancel mid-scan releases the reader.