Purpose

Queries like SELECT variant_get(v, '$.age', 'int') FROM T on a shredded Variant column still read all sub-columns and reassemble the full binary Variant, leaving Paimon's VariantRowType / clipVariantType infrastructure unused.

This PR adds PushDownVariantExtract (Spark 4 only), a Catalyst optimizer rule that replaces VariantGet with GetStructField and sets variantProjections on PaimonScan, so only the accessed typed_value.* Parquet sub-columns are read.

The rule runs in the "User Provided Optimizers" batch (via experimentalMethods.extraOptimizations) to ensure it fires after V2ScanRelationPushDown has built the scan relation.

Part of #4471

Note: Spark 4.0 lacks a V2-compatible variant push-down interface (SupportsPushDownVariantExtractions was introduced in 4.1), so registering a custom optimizer rule via experimentalMethods.extraOptimizations is the right fit for 4.0. For a future paimon-spark-4.1 module, a cleaner approach would be implementing SupportsPushDownVariantExtractions on PaimonScan and letting Spark's built-in V2ScanRelationPushDown handle the rewrite natively.

Tests

VariantTest.scala::VariantPushDownPlanTest (paimon-spark-4.0)