feat: size-balanced file assignment in Partitioner

[cigrainger]

Summary

• Replace round-robin with greedy first-fit-decreasing bin-packing for distributing Parquet files across workers. Files are sorted largest-first and assigned to the worker with the smallest current total load. O(n log k) time, within 11/9 OPT of optimal.
• Tiered file size fetching: File.stat for local files. Remote (S3/HTTP) files default to round-robin but support opt-in size balancing via parquet_file_metadata (reads Parquet footers for num_rows).
• S3 glob expansion on coordinator: S3/HTTP globs are now expanded via DuckDB's glob() (uses ListObjectsV2) on the coordinator, so individual files are distributed across workers instead of each worker expanding the full glob independently.
• Consolidates Parquet glob and DuckLake file distribution into a single distribute_parquet_files/5 codepath.

How it works

Before (round-robin):  [10GB, 1MB, 1MB, 1MB, 1MB, 1MB] across 3 workers
  Worker 1: [10GB, 1MB]  — massively overloaded
  Worker 2: [1MB, 1MB]
  Worker 3: [1MB, 1MB]

After (size-balanced):  same files, 3 workers
  Worker 1: [10GB]        — big file gets its own worker
  Worker 2: [1MB, 1MB, 1MB]
  Worker 3: [1MB, 1MB]

Test plan

• Skewed file sizes produce balanced assignment (big file isolated)
• Balance ratio < 4x on synthetic skewed data (5000/3000/1000/100/100/100 row files)
• S3 paths expand via glob and fall back gracefully
• All existing partitioner, DuckLake, coordinator, and distributed tests pass
• 578 tests, 0 failures, Credo clean

🤖 Generated with Claude Code