feat: add S3/Iceberg sink (append-only changelog)

[dariomazzitellireplik-coder]

Summary

New IcebergSink behind the opt-in cargo feature sink-iceberg (not in default builds). Streams CDC into Apache Iceberg tables on S3-compatible stores via an Iceberg REST catalog (Lakekeeper, Polaris, Nessie, Tabular, Glue REST endpoint).

Replaces the abandoned feat/s3-iceberg-sink branch, rebuilt from scratch after review found its write path unsalvageable (no-op catalog, deferred commits past LSN confirmation, deletes written as live rows, Parquet without Iceberg field-ids).

Design

• Append-only changelog: every INSERT/UPDATE/DELETE becomes one row with the source columns plus _cdc_op / _cdc_position / _cdc_ts metadata columns. Consumers materialize current state (dedup by PK + max _cdc_position, drop D).
• Durability before confirmation: each write_batch() writes Parquet via iceberg-rust's DataFileWriter (spec field-ids + column stats) and commits a FastAppend snapshot per touched table before returning — the pipeline's LSN confirmation never outruns durable data. No staging window; uncommitted files from a crash are simply unreferenced.
• Strict, exhaustive type mapping: UInt64 → decimal(20,0) (full range, no wrap), NUMERIC → decimal(p,s), UUID → uuid, temporal types to their Iceberg counterparts. Text-encoded values (snapshot path) are parsed strictly; unparseable/mismatched values fail the batch — never coerced to defaults.
• Honest capabilities: supports_upsert: false, supports_schema_evolution: false (iceberg-rust 0.9 transactions cannot update schemas — SchemaChange events feed the existing schema_evolution_skipped metric), optimal_flush_interval_ms: 60s to bound snapshot count.
• Tables auto-created (V2, unpartitioned, flat schema__table naming inside the SINK_DATABASE namespace); existing tables validated for column compatibility.
• REST catalog wrapper avoids the spec's HEAD endpoints (some servers, e.g. tabulario/iceberg-rest, reject them with 400) and handles this client version's 404 error mapping.

Also in this PR

• arrow/parquet 53 → 57 (required by iceberg 0.9). The Snowflake sink — the other arrow consumer — compiles and passes its test suite against v57.
• docs/configuration.md: Iceberg env var reference + operator requirements (snapshot expiry/compaction is operator-owned).
• Per-sink README: src/connectors/sinks/iceberg/README.md.
• Version 2.5.0 → 2.6.0, CHANGELOG under [Unreleased].

Verification

• cargo test --features sink-iceberg: 183 passed (includes Snowflake regression on arrow 57).
• cargo check on default, --features sink-iceberg, and --no-default-features --features 'sink-postgres sink-iceberg' — default build never references iceberg crates.
• cargo fmt --check + clippy --all-targets -- -D warnings clean.
• End-to-end against MinIO + tabulario/iceberg-rest + PG16: initial snapshot (3 rows) + live INSERT/UPDATE/DELETE each committed as Iceberg snapshots; read back exactly with pyiceberg 0.11 (field-ids, decimals, UUIDs, NULLs, timestamptz, op ordering by _cdc_position).
• Crash semantics: SIGKILL + DML while down + restart → resumed from checkpoint, zero loss, only at-least-once duplicate I rows from the engine's re-snapshot; materialized state matched the source exactly.

v1 non-goals (documented)

Merge-on-read/equality deletes, schema evolution auto-apply, partitioned tables, non-REST catalogs, table maintenance (expiry/compaction).