Preserve source columns across Iceberg partition spec changes

[osipovartem]

Summary

• Read historical Iceberg partition source columns from Parquet when live data files use a different partition spec, instead of substituting current-spec partition tuples.
• Match equality/position deletes by partition spec ID and tuple; apply unpartitioned equality deletes globally with bounded plan fan-out.
• Keep CHANGES scans correct for historical deleted-only manifests and restrict equality-delete metrics pruning to equality keys.

Verification

• Independent read-only review: approved.
• Targeted lib tests: 110 passed, 2 ignored.
• Equality-delete integration: 2 passed; position-delete integration: 3 passed.
• Targeted all-targets Clippy passed (two pre-existing lint categories suppressed).
• Debug-build grouping benchmark, 10k files / 20 plans: 846 ms without global deletes, 911 ms with four global deletes (+7.7% for grouping only, not end-to-end query latency).

Live Snowflake/Horizon read and COW validation follows in Embucket/rustice#700.