GH-45948: [C++][Parquet] Variant shredding

[qzyu999]

PR Stack (merge in order):

1. GH-45946: [C++][Parquet] Variant decoding #50121 Variant decoding
2. GH-45947 : [C++][Parquet] Variant encoding #50122 Variant encoding
3. GH-45948: [C++][Parquet] Variant shredding #50232 ← YOU ARE HERE Variant shredding (this PR, depends on GH-45946: [C++][Parquet] Variant decoding #50121 and GH-45947 : [C++][Parquet] Variant encoding #50122)

All three PRs are part of the GH-45937 umbrella (Add variant support to C++ Parquet).

Rationale for this change

Implements variant shredding/unshredding for C++ (GH-45948), part of the GH-45937 umbrella. Enables decomposing variant binary columns into native typed Arrow columns for Parquet statistics-based predicate pushdown. Depends on #50121 (decoding) and #50122 (encoding).

Note: This PR depends on #50121 and #50122. Please review/merge those first.

What changes are included in this PR?

Adds variant_shredding.h / variant_shredding.cc implementing:

• VariantShreddingSchema — tree structure defining shredding targets (Primitive,
  Object, Array). C++ equivalent of Rust's ShreddedSchemaBuilder.
• IsVariantCompatibleWithType() — strict type compatibility with safe int widening,
  Float→Double widening, timestamp unit+timezone matching, and decimal scale matching.
• ShredVariantColumn() — column-level shredding producing {metadata, value, typed_value}.
  Template-refactored loops (ShredPrimitiveLoop<>, ShredBinaryLoop<>) for all
  15+ supported Arrow target types.
• ReconstructVariantColumn() — column-level reconstruction reassembling shredded
  columns back to variant binary. Supports all list-like typed_value types (List,
  LargeList, FixedSizeList, ListView, LargeListView).

Extends VariantBuilder with 3 methods for shredding support:

• BuildWithoutMeta() — produce value bytes without metadata (for primitives)
• UnsafeAppendEncoded() — zero-copy raw byte append
• SetAllowDuplicates(true) — last-value-wins dedup for reconstruction safety

Supported shredding targets (Rust parity):
Bool, Int8, Int16, Int32, Int64, Float, Double, String, LargeString, StringView,
Binary, LargeBinary, BinaryView, Date32, Timestamp(Micro/Nano, TZ/NTZ), Time64(Micro),
FixedSizeBinary(16) (UUID), Decimal128 (scale-matched)

Variant::Null semantics (Rust parity): Variant::Null (0x00) is stored in the
value column, NOT the typed_value column. Distinguishes variant-null from SQL NULL.

NullBuffer output (Rust parity): Optional out_null_bitmap parameter on
ReconstructVariantColumn for SQL NULL disambiguation (bit=0 where both value and
typed_value are null).

Known gaps (documented TODOs for follow-up PRs):

• Recursive Object/Array sub-schema shredding in object fields (primitives only currently)
• CastOptions cross-type coercion (Uint, Float16, Decimal32/64, TimestampSecond/Milli)
• FixedSizeList/ListView as shredding output targets (reconstruction accepts all)
• Value-absent schemas ({metadata, typed_value} without value)
• DECIMAL256 shredding target (compatibility check exists but shred/reconstruct not wired)

Are these changes tested?

335 total tests (114 new shredding + 221 prior) pass with BUILD_WARNING_LEVEL=CHECKIN
covering: schema definition, type compatibility, primitive round-trip for all supported
types, object shredding (full/partial/fallback), array shredding (recursive elements),
typed round-trip (Decimal128, UUID, all timestamps, Float→Double, Int8/Int16, LargeString,
LargeBinary, StringView, BinaryView), all list-like reconstruction, error cases, and
NullBitmap semantics.

Are there any user-facing changes?

New public API in arrow/extension/variant_shredding.h: VariantShreddingSchema,
IsVariantCompatibleWithType(), ShredVariantColumn(), ReconstructVariantColumn().
New methods on VariantBuilder: BuildWithoutMeta(), UnsafeAppendEncoded(),
SetAllowDuplicates().

AI Disclosure: AI coding assistants were used during development for scaffolding,
test generation, and review iteration. All code has been reviewed, debugged, and
verified by the author who owns and understands the changes.

• GitHub Issue: GH-45937: [C++][Parquet] Variant shredding #45948

[qzyu999]

Force-pushed the shredding implementation built on top of the refactored decoding/encoding layers. This uses the idiomatic C++ view classes and RAII builder from the parent PRs.

Highlights:

• Template-refactored shredding loops (ShredPrimitiveLoop<>, ShredBinaryLoop<>) eliminate per-type code duplication
• Recursive array element shredding (Rust parity)
• All 5 list-like types supported in reconstruction (List, LargeList, FixedSizeList, ListView, LargeListView)
• out_null_bitmap parameter for SQL NULL disambiguation (Rust NullBuffer parity)
• Metadata caching in reconstruction path (avoids redundant DecodeMetadata per row)
• Object sub-field native extraction (primitives only; Object/Array sub-schemas deferred)

335 tests pass with BUILD_WARNING_LEVEL=CHECKIN (114 shredding-specific).