feat(java): support json extraction by scanning

[majin1102]

Close #5744

For now arrow-java does't provide a convenient way like pa.json(). So we extend the field type in lance java binding to support related json udf

[github-actions]

Code Review for PR #5770

Summary

This PR adds Java utilities for creating Arrow JSON extension fields and a test to verify JSON extraction via scanning works correctly.

P0 Issues

None identified - The implementation is correct and addresses the issue #5744.

P1 Issues / Suggestions

1. Missing Javadoc on constants (Minor)
   The class uses EXTENSION_NAME_KEY and ARROW_JSON_EXTENSION_NAME constants but doesn't document their relationship to the Arrow specification. Consider adding a brief class-level Javadoc explaining that these constants align with the Arrow JSON extension type specification (arrow.json).

2. Consider adding jsonLargeBinary helper
   Looking at rust/lance-arrow/src/json.rs, Lance internally uses LargeBinary with lance.json extension for JSONB storage. The current PR only exposes arrow.json extension (which uses Utf8/LargeUtf8 and gets converted to JSONB on write). This is the correct approach for users writing JSON data, but a comment noting this would help future maintainers understand the data flow:

   • User writes arrow.json (Utf8) → Lance converts to lance.json (LargeBinary JSONB) on write
   • On read, Lance converts back to arrow.json (Utf8)

3. Test coverage is good - The test properly validates:

   • Creating a dataset with JSON extension field
   • Writing JSON data
   • Querying with json_extract filter
   • Verifying correct row filtering

Positive Aspects

• Correct use of Collections.unmodifiableMap for immutable metadata
• Both Utf8 and LargeUtf8 variants are provided
• Test follows existing patterns in the codebase (uses @TempDir, proper resource cleanup with try-with-resources)
• The solution aligns with how PyArrow's pa.json_() works, providing parity between Python and Java SDKs

Verdict

LGTM ✓ - The PR correctly solves the issue and the implementation is clean. The test adequately covers the use case.

[majin1102]

@Xuanwo Could you take a look at this json related feat for Java binding?

Thanks