feat: parse logical types into semantic types and output encodings

[Xuanwo]

Part of implementing #7073 (spec in #9485). This adds the internal model the semantic type contract needs and routes schema compatibility checks through one entry point, without changing the behavior of any existing table.

LogicalType::semantic() reads a logical_type string as a semantic type (string, binary, list, decimal:<p>:<s>, json, blob, or an unchanged type) plus the output encoding a legacy alias implies: large_string is string with large_utf8, dict:string:int16:false is string with dictionary:int16:utf8, and decimal:256:10:2 is decimal:10:2 with decimal256. OutputEncoding parses and prints the values of the lance-schema:output-encoding field entry and checks them against a semantic type. Schema::to_canonical_types() produces the form a table under the contract records, keeping field IDs, nullability, metadata, and children; nothing calls it yet. The model stays internal to the Rust core; there is no plugin trait.

Compatibility checks now decide field types through Field::type_matches with a TypeComparison. SchemaCompareOptions carries it, and every check against a dataset takes it from Dataset::type_comparison(), as does the Merge commit check for field bindings. Every table currently compares exactly, so Field::data_type(), compatibility results, and error messages stay the same. The follow-up derives TypeComparison::Semantic from FLAG_SEMANTIC_TYPES.

The only parsing change is that the canonical decimal:<p>:<s> now parses, to Decimal128 when p ≤ 38 and Decimal256 otherwise. Nothing writes it yet.

Checks that compare Arrow DataTypes directly (project_by_field, do_intersection, Field::merge, the partial-schema check in merge_insert, index copying and index segment merges) are left unchanged. They compare physical layouts on both sides, and the write path in the follow-up decides how they apply to tables with mixed layouts.