feat: support milvus backfill with milvus snapshot

[shaoting-huang]

The backfill feature allows adding new fields to existing Milvus collections by reading collection data, joining it with new field data, and writing binlog files directly to S3.

Core Components

• MilvusBackfill.scala - Main orchestrator (src/main/scala/operations/backfill/)
• BackfillConfig.scala - Configuration (Milvus connection, S3 settings)
• SegmentPartitioner.scala - Custom Spark partitioner for segment-aware distribution

Data Flow

1. Read metadata - From Milvus snapshot (preferred) or client API
2. Load data - Read backfill Parquet (pk, new_fields...) + collection data (pk, segment_id, row_offset)
3. Join - Sort merge join on primary key to match new data with original rows
4. Partition by segment - Use SegmentPartitioner for 1:1 segment-to-Spark-partition mapping
5. Write - Each executor writes binlog files to S3 via MilvusLoonWriter (Storage V2 FFI)

Key Features

• Snapshot mode - Zero-client offline backfill using MilvusSnapshotReader (Milvus 2.6+)
• Segment preservation - Maintains original segment structure with segment_id and row_offset
• NULL handling - Supports partial backfill (unmatched PKs get NULL values)
• Storage V2 - Direct FFI/Loon writer for efficient binlog generation

Output

Writes to: s3://{bucket}/{root_path}/insert_log/{collectionID}/{partitionID}/{segmentID}/new_field/

Returns Either[BackfillError, BackfillResult] with per-segment row counts and manifest paths.