Add equality-delete reads and atomic upserts

[nicosuave]

Summary

• index equality deletes during scan planning with sequence, partition, schema-evolution, null, and bounds semantics
• apply equality deletes to Parquet and ORC Arrow scans with field-ID alignment, hidden key projection, and record-batch anti-joins
• write Iceberg v2 equality-delete files and commit delete/data manifests in one row-delta snapshot
• add Table.upsert_by_equality_delete, including delete-only keys for atomic hard deletes
• encode equality field IDs with the Iceberg-compatible Avro int wire type

This is a Sidequery fork review PR. It does not open or modify an Apache pull request. The implementation incorporates correctness cases discussed around apache/iceberg-python#3805 while adding the write path needed by dlt-iceberg.

Semantics and current limits

• Snapshot visibility is atomic; physical files written before a failed metadata commit may require orphan-file cleanup.
• Concurrent commits are ordered by snapshot sequence: later equality deletes can hide earlier matching rows, while later data survives.
• Writes require Iceberg v2+, reject nested equality fields, and require partition source columns in the equality key.
• Evolved partition specs are rejected for equality-delete upserts.
• Reads anti-join per record batch, but equality-delete files are still cached in memory for a scan; this reduces data-file amplification but is not yet a spilling execution backend.

Validation

• full hooks: passed
• unit suite before the final delete-only API addition: 3,970 passed, 3 skipped
• focused upsert suite after the final addition: 29 passed
• local composite merge: 15,000 source rows passed
• Docker REST/MinIO/Spark interoperability: 200,000 source rows, 300,000 final rows passed
• additional local stress proof: 750,000 source rows into a 1,000,000-row table, 1,250,000 final rows passed