datalake_fdw: storage I/O over object storage (S3) #2009

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

Scope

• An S3 backend (AWS SDK for C++) in backend_registry, configured from the volume server and user-mapping options the validator already accepts (endpoint, region, path style, credentials, session token).
• Arrow RandomAccessFile / OutputStream adapters over common/file_system_wrapper.h; parquet_read.cpp and parquet_write.cpp are the only format files that change.
• The writer's rule carried over: an object is created once, never overwritten, and a failed writer removes exactly that object.
• A storage_s3 smoke category against MinIO, and the CI job that starts MinIO (the CATEGORY_SERVICES hook in run_smoke_tests.sh).

Out of scope

HDFS (libhdfs3 is unmaintained; the wrapper leaves room for a second backend). Iceberg path layout (C).

Depends on

Nothing. Starts first.

Acceptance

• The datalake_fdw: read and write Parquet through Arrow #1951 type set round-trips through s3:// against MinIO in CI.
• A write aborted after the first row group leaves no object; an existing key is refused.
• Credentials never appear in logs, errors or EXPLAIN.
• nm -D exports only the PostgreSQL entry points; Arrow 9 and 17 build.