feat: add position_deletes metadata table

[wirybeaver]

Summary

Add general position_deletes metadata inspection for Parquet delete files and Puffin DVs, independent of the high-level delete-update API.

• preserve stored row payloads, subset-row nulls, and partition evolution with collision-free metadata field IDs

• support mixed upgraded tables, multi-blob DVs, empty scans, and complete top-level projections

• adapt to the current metadata-table ArrowArrayStream API, emitting bounded batches

• detect optional physical Parquet row fields through a reader schema-presence query, not error-message matching

• construct and validate projection state once per scan, reusing it across batches; open each position-delete file once while still distinguishing absent row columns from invalid null rows

• fix decimal partition appending/decoding and honor precision-dependent Avro fixed widths in both encoder paths; preserve logical date values represented by primitive storage literals

DV-PR-3 in #916. Based directly on main at 3c5715c; no dependency on #913, #914, or #915. Reuse the existing range-based DV reader without full footer cross-validation.

Test plan

GCC 14 focused checks on the four-feature integration tree based on 3c5715c; the final changed branch sources were verified to match the tested tree.

• position_deletes_table_test: 19 passed
• manifest_test --gtest_filter='*ManifestWriter*:*ManifestReader*': 72 passed, 7 version-specific skips
• avro_test --gtest_filter='*Decimal*': 6 passed
• data_test --gtest_filter='ArrowRowBuilderTest.*': 8 passed

Coverage includes decimal/date partition write/read, schema-sized signed decimal encoding, reordered projection across multiple batches, optional/subset stored rows, and empty/stale-layout scans.

• Repository pre-commit checks and git diff --check: passed

Deliberate limits

Results are materialized before returning the stream; bounded batches are not a claim of lazy or memory-bounded input scanning. Nested projection such as row.id remains outside this feature and is rejected explicitly.

Recreate the inspector after source schema/format/partition-spec changes; scans reject a stale layout rather than mixing schemas.