Optimize partition-only deletes

[jdctinuiti]

Summary

This PR adds a metadata-only DELETE path for DuckLake tables when the DELETE can be proven to remove whole data files, and fixes table-statistics accounting when whole files are dropped.

Metadata-only partition DELETE

For DELETE predicates over identity partition columns, DuckLake can retire matching data files by updating metadata instead of scanning files and writing row-level delete files.

This path is used only when all candidate files are safe to drop:

• the predicate is reducible to accepted constant values for identity partition columns (= / IN)
• every candidate file has recorded partition values matching the accepted set
• no candidate file has active row-level delete files
• no inlined data or active inlined deletion table is involved
• non-identity partition transforms, mixed non-partition predicates, NULL partition predicates, and legacy files without partition metadata fall back to the existing row-level delete path

Full-table DELETE on partitioned tables can also use this path when the scan selects whole active data files.

Stats fix

Whole-file drops now decrement ducklake_table_stats.record_count and file_size_bytes.

Before this PR, full-file deletes retired data files but left global table stats unchanged, so repeated delete/reinsert cycles inflated:

• optimizer cardinality estimates via GetCardinality
• duckdb_tables size reporting

next_row_id remains monotonic. Column min/max stats are left unchanged because exact recomputation would require scanning remaining files.

The stats cache is also invalidated after delete-only commits that drop files, since next_file_id does not change for those commits.

Tests

Adds coverage for:

• metadata-only partition DELETE with = and IN
• intersecting partition filters
• fallback for mixed predicates, NULL predicates, transformed partitions, legacy files, existing deletes, and inlined data
• atomic delete + ducklake_add_data_files
• concurrent conflicting metadata deletes
• table-stats decrement and stats-cache invalidation

[kinghuang]

Nice, I've been looking for something like the efficiency of attaching and dropping table partitions in PostgreSQL.

[jdctinuiti]

Rebased this stack onto current origin/main and cleaned the branch history so the PRs carry only their feature changes. I intentionally left unrelated DuckDB-main compatibility fixes out of this PR stack; those should land separately. The only compatibility cleanup folded into #1171 is for expression accessors in the metadata-delete planner code introduced by this PR.