[format] Support nested variant column pruning for shredded parquet files

[juntaozhang]

Purpose

This PR introduces VariantShreddingTypePruner, which builds a path trie from the logical Variant projection and recursively prunes unneeded typed columns while preserving value fallbacks for paths that cannot be satisfied from typed columns. Supported cases include:

• Nested object field pruning, e.g. $.a.b.
• Nested fields inside array elements, e.g. $.arr[0].x.
• Top-level array element pruning, e.g. $[0].x.
• Case-insensitive field matching with duplicate-field detection.
• Canonical three-level Parquet lists.

Two-level or non-canonical list layouts and map-typed typed_value groups are kept unchanged as a safe fallback.

Tests

• VariantShreddingReadTest#testReadNestedVariantWithPruning: end-to-end read test for both shredded and unshredded Parquet file, covering nested object fields, array index paths, missing fields inside array elements, and reading whole array elements as VARIANT.
• VariantShreddingTypePrunerTest: unit tests covering
  • nested object field pruning,
  • missing top-level / nested field fallback to value,
  • mixed existing and missing fields,
  • whole-field keepAll semantics and so on.

[JingsongLi]

When this shredding node is object-typed but the projection continues with an array segment (for example, the schema defines a as an object, a row contains {"a":[{"x":1}]}, and the read projects $.a[0].x), node.children is empty, so needValue remains false and this method produces an empty nested shredding row. The writer stored the type-mismatched array in a.value, so dropping it makes the scan fail with Invalid variant shredding schema: ROW<> NOT NULL; the same case succeeds on the base revision. Please retain value whenever node.arrayElement != null (and apply the symmetric fallback in the list branch for object children), with an end-to-end heterogeneous Variant test.

[juntaozhang]

Thanks for pointing out this issue, I have fixed it. PTAL

[JingsongLi]

Variant object keys are resolved exactly downstream through objectSchemaMap.get(objExtr.getKey()), so lowercasing only in the pruner can select typed key a for requested path $.A and then drop the raw value that actually contains the exact key A. With case-insensitive table mode, shredding schema key a, row {"A":7}, and projection $.A, the base revision returns 7 but this change returns null. Please keep Variant-path key matching case-sensitive independently of Parquet column-name resolution, or make the entire extraction and fallback pipeline use one consistent semantic.

[juntaozhang]

Thanks for the guidance, I have removed the case-insensitivity support. PTAL

[JingsongLi]

[P1] Preserve the fallback for mixed array/object projections

This branch selects list pruning whenever an array projection is present, even if the same Variant column also has object projections in node.children. For a list-typed shredding schema, querying $[0].x and $.y together therefore removes the parent value field; an object-shaped row such as {"y": 2} is stored only in that fallback and $.y silently returns null. I reproduced this on this head, and adding node.children.isEmpty() to this condition makes the regression test pass. Please either skip list pruning when object children coexist or retain value in that case.