refactor: read JSON through one accessor in the JSON semantic module

[Xuanwo]

Part of #7073. Builds on #9478.

Problem

Code that computes on JSON decided whether a column holds JSON by matching lance.json over LargeBinary itself. The same JSON column was therefore tokenized as JSON when it came from storage but as plain text when it arrived as Arrow JSON text (arrow.json over Utf8), and the JSON scalar index could not read Arrow JSON text at all. JSONPath selection was also written twice: JsonArray::json_path kept only the first match, while json_extract returns every match as one JSON array.

Behavior

• lance_arrow::json is the one place that encodes and decodes JSONB and selects JSONPath values. JsonEncoding::of_field recognizes a JSON field in either representation, and JsonValues reads its values as JSONB or as text, whichever the consumer needs, regardless of the representation they arrived in.
• FTS index builds and unindexed FTS input read JSON documents through JsonValues, so both representations are tokenized as JSON. The JSON scalar index accepts Arrow JSON text and indexes the same values it would from JSONB.
• The SQL JSON functions use the shared JSONPath selection. JsonArray::json_path now returns every match of a multi-match path as one JSON array, the same as json_extract.
• Breaking API changes: lance_index::scalar::inverted::json::JsonTextStream::new is replaced by the fallible try_new, which rejects a column that does not hold JSON, and jsonb_to_json is removed in favor of JsonValues::to_text.
• No format change, and blob behavior is unchanged. JsonValues is an internal Rust API, not a plugin extension point.

Tradeoffs

• dataset/sql.rs, lance-datafusion/src/projection.rs and io/exec/projection.rs do not inspect JSON metadata themselves: SQL only converts its output to Arrow JSON through SchemaAdapter, and the projections carry field metadata through unchanged. They are left as they are; choosing the user-facing output encoding belongs to the read path.
• An FTS index records its tokenizer when it is created, so existing indices keep theirs. The document-type change only affects new indices built from Arrow JSON text.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The shared JSON accessor is preferable to per-consumer metadata branching: FTS, JSON scalar indexing, and JSONPath evaluation apply the same representation semantics to stored JSONB and Arrow JSON text, while existing index tokenizer metadata remains intentionally preserved. The affected accessor, UDF, and index paths have focused regression coverage, including multi-match JSONPath selection.