feat(python): pass merge_insert sources as scannables

Python's `execute_merge_insert` now takes a `PyScannable` (like `add()`),
and the source collection threshold is exposed as a
`source_collect_threshold_bytes` kwarg on `execute()` and a
`merge_insert_source_collect_threshold_bytes` connection option.

Node passes the decoded IPC batches as a `Vec<RecordBatch>`, so the
merge join sees their exact row count.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

nodejs/__test__/table.test.ts

@@ -1081,6 +1081,16 @@ describe("merge insert", () => {
     expect(res).toEqual(expected);
   });
 
+  test("empty source with delete by source", async () => {
+    const empty = arrow.makeEmptyTable(await table.schema());
+    const res = await table
+      .mergeInsert("a")
+      .whenNotMatchedBySourceDelete()
+      .execute(empty);
+    expect(res.numDeletedRows).toBe(3);
+    expect(await table.countRows()).toBe(0);
+  });
+
   test("timeout", async () => {
     const newData = [
       { a: 2, b: "x" },