[core] Restore shredding write plans from historical files (variant & map shredding)

[lxy-9602]

Purpose

This is a subtask of MAP shared-shredding support.

This PR introduces a shared history mechanism for restoring per-file shredding write plans from recent data files.

• Add format-independent ShreddingFileMetadata, ShreddingWritePlanHistory, and ShreddingWritePlanPreparation abstractions.
• Lazily load physical schemas and field metadata from recent restored files, while preserving schema evolution and column renames through field IDs.
• Update history after each rolling file so subsequent files can use the latest metadata.
• Restore MAP shared-shredding column counts from historical maximum row widths. Files without valid history continue to use the configured max-columns.
• Add the opt-in variant.restoreShreddingSchemaFromHistory option. Existing Variant inference behavior remains unchanged by default; missing Variant fields still fall back to row sampling.
• Propagate history through append-only, primary-key, changelog, thin-mode, and compaction/rewrite writers. Rewrite outputs derive history from their actual input files.
• Support metadata recovery through both ORC and Parquet, including mixed-format history.

Tests

Over 1,900 lines of this PR are dedicated to comprehensive test coverage.

Added or updated coverage for:

• Shredding history ordering, file limit, schema-ID normalization, column renames, and recoverable metadata IO failures.
• MAP adaptive column-count restoration, multiple independent MAP fields, rolling files, mixed ORC/Parquet history, and partition/bucket isolation.
• Variant schema restoration, nested Variant fields, partial restoration, and fallback inference for newly added Variant fields.
• Append-only and primary-key tables, including thin mode, aggregation, changelog files, external spill, deletion vectors, and regular/partial/clustering compaction.
• ORC and Parquet metadata round trips and per-file history updates.