Add VariantExprAdapterFactory to read files with a different shredding layout - #149

Adds VariantExprAdapterFactory, a PhysicalExprAdapterFactory that wraps DataFusion's default adapter. When a Variant column's layout in a Parquet file differs from the merged table layout, the default adapter casts the file struct to the table struct and null-fills the missing typed_value children, which the shredding spec reads as "not an object"; every shredded field of that file then returns NULL with no error. The new adapter instead reads the file column as is and, per batch, unshreds it and shreds it again to the table's typed_value schema. Files whose layout already matches the table are left to the default adapter.

Closes #148.

Installed per table with ListingTableConfig::with_expr_adapter_factory(Arc::new(VariantExprAdapterFactory::default())); the README has the full snippet. Wiring it in automatically for the extension type is a core question, raised on apache/datafusion#27089.

Two limits worth knowing: files that need the rewrite read the whole Variant column, so they don't get leaf-level projection; and only whole-column reads of the Variant are rewritten, so direct struct access like event['typed_value'] on a mismatched file still sees the null-filled cast.

Tests write real Parquet files and read them through a ListingTable: a shredded plus an unshredded file (3 of 6 rows NULL without the adapter, 6 of 6 with it), two files shredded on different fields, a whole-column read, and same-layout files showing no rewrite in the plan.

Validation: cargo test (new and existing tests, SLTs, README examples), cargo fmt --check, cargo clippy --all-targets -- -D warnings.