feat: add count_table_rows, insert_into_table, query_table for DirectoryNamespace

[XuQianJin-Stars]

close #6111

Add count_table_rows, insert_into_table, query_table for DirectoryNamespace

• Enable lance-io 'tencent' feature in java/lance-jni/Cargo.toml for COS support
• Implement count_table_rows with version checkout and predicate filter
• Implement insert_into_table with append/overwrite modes via Arrow IPC
• Implement query_table with filter, limit, offset, version support
• Add 8 unit tests covering all new methods and edge cases

[github-actions]

PR Review

P0: query_table semantic mismatch with QueryTableRequest

The QueryTableRequest schema (per rest.yaml) has vector and k as required fields — it's designed for vector similarity search, not general-purpose table scanning. This implementation ignores the vector field entirely and repurposes k as a generic row limit for a plain scan. This will produce confusing behavior:

• Callers must provide a vector (it's required in the schema) even though it's unused.
• k means "number of nearest neighbors" in the API contract, not "LIMIT".
• Filter-only queries without vector search are not what this endpoint is designed for.

If the goal is to support simple table scans, consider implementing a different method or at minimum handle the vector field appropriately (e.g., perform actual vector search when a vector is provided, or return an error if no vector is given and the method doesn't support plain scans).

P1: Unnecessary memory copy in insert_into_table

let cursor = Cursor::new(request_data.to_vec());

request_data.to_vec() copies the entire IPC payload. Bytes implements AsRef<[u8]>, so you can use Cursor::new(request_data.as_ref()) or Cursor::new(&request_data[..]) to avoid the allocation.

P1: query_table collects all scan results into memory then re-serializes

let batches: Vec<_> = scanner.try_into_stream().await...try_collect().await...;

For large tables this buffers everything in memory twice (once as RecordBatches, once as the IPC buffer). Consider streaming batches directly to the IPC writer instead of collecting first.

Minor: Unrelated feature flag change

Enabling the tencent feature in java/lance-jni/Cargo.toml is bundled into this PR but is unrelated to the DirectoryNamespace method implementations. Consider separating it into its own PR to keep changes focused.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Xuanwo]

Hi, thank you @XuQianJin-Stars for working on this. I noticed that this PR combines multiple different goals into one, which makes it difficult to review and cherry-pick. Could you split them into separate PRs instead?

[XuQianJin-Stars]

Hi, thank you @XuQianJin-Stars for working on this. I noticed that this PR combines multiple different goals into one, which makes it difficult to review and cherry-pick. Could you split them into separate PRs instead?

separate PRs：
#6131
#6132