feat: Add typed JSONPath extraction functions

Problem

Lance currently has two different JSON access patterns:

• json_get_string, json_get_int, json_get_float, and json_get_bool accept one literal object key or array index and return a typed scalar.
• json_extract accepts JSONPath, but returns the extracted value as serialized JSON text.

As a result, reading a typed nested scalar requires composing json_get calls:

json_get_string(json_get(json_get(data, 'user'), 'profile'), 'name')

Passing a dotted path such as json_get_string(data, 'user.profile.name') cannot safely be reinterpreted as JSONPath because JSON object keys may legitimately contain dots.

Proposed contract

Add typed JSONPath extraction functions, for example:

json_extract_string(data, '$.user.profile.name')
json_extract_int(data, '$.user.age')
json_extract_float(data, '$.metrics.score')
json_extract_bool(data, '$.flags.active')

The exact names are open to design, but the API should:

• use the same JSONPath syntax and evaluator as json_extract;
• return typed Arrow scalar values;
• preserve the current literal-key semantics of json_get_*;
• define behavior for a missing path, JSON null, incompatible types, nested objects, and arrays;
• define whether conversions match the existing json_get_* coercion rules.

Acceptance criteria

• Typed extraction works for nested object fields and array elements.
• Missing paths and JSON null have documented SQL null behavior.
• Type mismatches and coercions are consistent and covered by tests.
• A literal key containing . remains accessible through json_get_* without ambiguity.
• DataFusion UDF tests and Python SQL/filter integration tests cover the new functions.
• User-facing JSON query documentation explains when to use json_get_*, json_extract, and the typed JSONPath functions.

Related work

• JSON support epic: https://github.com/lance-format/lance/issues/4516

Start by locating the existing json_get_* and json_extract implementations, the DataFusion UDF tests, and the Python SQL/filter integration tests. Compare their JSONPath evaluation, scalar typing, null handling, and coercion behavior before choosing the function names and contract. Done means nested objects and arrays work, literal dotted keys remain supported, documented SQL null and mismatch behavior is covered by tests, and the user-facing JSON query documentation is updated.