[wip]feat: support column default values (initial-default / write-default)

[huan233usc]

Closes #730 (item 2 of #637).

Implements Iceberg v3 column default values: initial-default / write-default on the
schema, JSON serde, read-path application, schema-evolution support, and format-version
validation. Semantics follow the Java reference implementation
(Types.NestedField / SchemaParser / SchemaUpdate / Schema.checkCompatibility).

What changed

Schema model

• SchemaField carries initial_default / write_default literals, stored as
  std::shared_ptr<const Literal> — an immutable payload shared across field copies,
  the same pattern as the adjacent type_ member (and the C++ analog of Java's
  final Literal<?> reference). literal.h cannot be included from schema_field.h
  due to the literal.h -> type.h -> schema_field.h include cycle, and sharing keeps
  SchemaField copies cheap. Accessors return
  std::optional<std::reference_wrapper<const Literal>>, the same optional-reference
  idiom as Schema::FindFieldByName, mirroring Java's nullable Literal<?> on
  Types.NestedField; WithInitialDefault / WithWriteDefault copy-modifiers follow
  the style of AsRequired/AsOptional.
• SchemaField::Validate() checks that defaults are primitive literals matching the
  field type (Java: castDefault rejects defaults on nested types);
  Schema::Validate(format_version) rejects initial-default below v3 with the same
  message wording as Java's Schema.checkCompatibility, using the previously-unused
  TableMetadata::kMinFormatVersionDefaultValues (resolves the TODO there).

JSON serde

• FieldFromJson / ToJson(SchemaField) parse and write initial-default /
  write-default using the existing single-value serialization
  (LiteralFromJson(json, type)), mirroring Java's SchemaParser +
  SingleValueParser, and resolving the add default values TODO in struct
  serialization. All primitive types supported (incl. decimal, fixed, uuid, temporal).

Read path (initial-default)

• Project() maps a column missing from a data file to
  FieldProjection::Kind::kDefault carrying the literal when the field has an
  initial-default — for required and optional fields, per the spec. All four
  projection/decoding paths are covered: schema_util.cc (resolving its default-value
  TODO), parquet_schema_util.cc (used by the Parquet reader), avro_schema_util.cc,
  and the Avro direct decoder.
• New iceberg::arrow helpers (literal_util) convert a Literal to an Arrow scalar /
  constant array; the Parquet reader materializes kDefault via MakeDefaultArray and
  the Avro reader via AppendDefaultToBuilder.

Schema evolution (write-default)

• AddColumn / AddRequiredColumn accept an optional default_value, used as both the
  initial-default and write-default of the new column. Like Java's
  Types.NestedField.castDefault, the value is cast to the column type rather than
  rejected (failing only if it cannot be cast). A required column with a default does
  not need AllowIncompatibleChanges().
• RequireColumn() accepts a column added with a default in the same update (resolves
  the defaulted-add TODO in UpdateColumnRequirementInternal).
• New UpdateColumnDefault() updates the write-default of an existing column
  (initial-default stays fixed once the column exists), matching Java's
  updateColumnDefault.
• UpdateColumnDoc / RenameColumn / UpdateColumn and the ApplyChangesVisitor
  preserve defaults when reconstructing fields; type promotion casts the defaults to
  the new type.

Scope note: write-path application

Writers in this library consume complete Arrow arrays, so filling omitted columns with
write-default at write time remains the engine's responsibility, as in Java. The
library's role — storing, validating, serializing the defaults, and exposing them
through schema evolution — is covered here.

Testing

• literal_util: dedicated unit tests converting every primitive type to Arrow
  scalars (incl. negative decimals, uuid, all timestamp variants), constant-array
  materialization, builder append, and rejection of null/sentinel literals.
• Defaults are validated to be non-null, non-sentinel values: Literal::CastTo
  signals narrowing with AboveMax/BelowMin sentinels, which are rejected when
  setting a default (e.g. a long default too large for an int column).
• Schema serde round-trips (top-level + nested struct fields, mismatch rejection).
• Schema::Validate: v2 rejects initial-default, v3 accepts; mismatched default
  type rejected.
• Projection: missing required/optional fields with initial-default -> kDefault;
  present fields ignore initial-default.
• Parquet ProjectRecordBatch and Avro AppendDatumToBuilder: missing columns
  materialize the default at top level and in nested structs.
• File-level end-to-end: write a Parquet/Avro file with the original schema, read it
  back through the real readers with an evolved schema carrying initial-defaults, and
  verify every row is filled (ReadMissingFieldsWithDefaults in both reader tests).
• UpdateSchema: add with default (both defaults set), required-with-default without
  AllowIncompatibleChanges(), uncastable default rejected, UpdateColumnDefault on
  added and pre-existing columns (incl. int->long cast), RequireColumn on defaulted
  add, doc-update preservation, type-promotion casting, and v2 rejection at Apply();
  new TableMetadataV3Valid.json test resource.
• Full suite passing locally (the pre-existing S3 file_io_test is unrelated and fails
  only in my local environment due to a Homebrew AWS-SDK ABI issue; not touched by this
  change).

This pull request and its description were written by Isaac.

[huan233usc]

Alternative to those with methods are make it in constructor