feat: migrate existing tables to the semantic type contract

[Xuanwo]

Do not merge before #9485, the format specification this implements, is voted and merged. Stacked on #9494.

Tables created before the semantic type contract of #7073, and every table on data storage 2.2 or earlier, stay legacy tables: they reject appends in another layout and cannot use output encodings. This adds the explicit upgrade the design calls for.

Dataset::migrate_to_semantic_types() commits one metadata-only version. It rewrites each legacy alias to its canonical type with the output encoding that keeps the current read type (large_string becomes string with large_utf8, decimal:256:10:2 becomes decimal:10:2 with decimal256), and sets FLAG_SEMANTIC_TYPES. No data file is rewritten. Reads return the same Arrow types, values, and field IDs, and appends then accept any layout of a column's type. It is idempotent, there is no downgrade, and restoring an earlier version keeps the flag. Python exposes LanceDataset.migrate_to_semantic_types() and Java Dataset.migrateToSemanticTypes().

The migration refuses, and commits nothing, in three cases:

• The table references data files of version 2.0 or earlier. Flagged tables need 2.1 or later.
• A column has no semantic type, meaning a dictionary whose values are not strings or bytes. The admission rule gives such a column no canonical form, so it cannot move without changing its read type.
• A field already carries a lance-schema:output-encoding entry. Legacy tables ignore it, so adopting it would change what reads return.

The commit follows migrate_to_stable_row_ids: it projects the current schema with a commit option that sets the flag, and it does not retry past a concurrent write. This keeps the transaction format unchanged.