VARIANT: targeted extract pushdown for Parquet, unshredded read crash fix, and unshredded write support

[fuziontech]

Summary

Path queries on VARIANT columns (variant_extract(v, 'field'), v.field with a cast or filter) on unshredded and partially-shredded Parquet files were decoding every row's entire value twice: the reader converted the full binary value into a variant vector (re-encoding all of it via BuildVariant), then the extract decoded it again to pick out one field. The metadata dictionary was additionally decoded — with all key strings allocated — once per row.

This PR makes extract pushdown do a targeted decode (the same approach ClickHouse took in fuziontech/clickhouse#1), fixes a crash on unshredded schemas, and adds spec-compliant unshredded writes.

Benchmarks (3M rows, aarch64, best-of-5)

Query	duckdb 1.5.5	ClickHouse (fuziontech/clickhouse#1)	this branch
path groupby shredded	1740ms	562ms	~20ms
path filter shredded	1733ms	507ms	~20ms
path groupby unshredded	1537ms	590ms	~110ms
path filter unshredded	1534ms	514ms	~110ms
full scan shredded	10545ms	~4000ms	~2400ms (decode-only ~1000ms)
full scan unshredded	10015ms	~3600ms	~2600ms (decode-only ~1100ms)
pruned groupby/filter (fully-shredded)	n/a	~71ms	~20ms

Changes

• Targeted extract in VariantColumnReader (parquet_variant_iterator.cpp, variant_column_reader.cpp): when the extract is pushed down, navigate only the requested path per row — binary values via the object field table, shredded objects via the typed tree with binary-overlay fallback — and emit just the addressed value. No whole-value conversion, no double pass. BY_KEY path components are resolved to metadata dictionary field ids once per distinct dictionary (keyed by blob content), so per-row lookups are integer comparisons.
• Metadata dictionary caching: ParquetVariantIterator::GetMetadata() decoded the key dictionary per row; it's now reused across rows sharing identical metadata bytes (writers emit identical metadata per row). The resolution cache is keyed by blob content, so per-row/per-row-group dictionary changes stay correct.
• sorted_strings is verified, not trusted: the flag is unreliable across writers (DuckDB itself set it without always sorting until f129a86); sortedness is checked on the decoded dictionary once per distinct dictionary instead.
• ParquetObjectIterator: skip std::sort when fields are already in lexicographic key order (the common, spec-compliant case).
• Crash/wrong-result fix for unshredded (2-child) schemas: the pushdown index was silently dropped for 2-child (metadata+value) variant schemas, so the scan produced full or wrongly typed vectors — Vector::Shred can only be used on variant vectors for extracts with a pushed-down cast. The 2-child case now goes through the same pushdown path as shredded schemas.
• Unshredded write support: SHREDDING {col: 'VARIANT'} now writes a spec-compliant unshredded (metadata+value) variant column instead of failing with Unsupported type "ANY"; the schema analyzer no longer overrides an explicitly requested layout. SHREDDING {col: 'NULL'} behavior is unchanged (analyzed partial shredding).

Testing

• New test/parquet/variant/pushdown/variant_extract_unshredded.test: unshredded writes (schema assertion: no typed_value), pushed-down casts, filters, missing keys, nested paths, row-group-local metadata dictionaries.
• test/parquet/variant/* (34 cases) and the full test/parquet/* suite (62 cases, 9164 assertions) pass.

Notes

• Companion DuckLake work (enabling this pushdown through DuckLake scans: the function.statistics optimizer blocker, inlined-data extracts, variant stats propagation) lives in the DuckLake repo and will come as a separate PR there. Verified end-to-end through DuckLake: path group-by/filter run in ~60-80ms for the same dataset.