WindowEvaluator already provides the pure-Python window UDF base class #1577 (item 6) asks for

Split out from #1577 (item 6), which asks for a `WindowUDF` ABC equivalent to `Accumulator`.

Checked this against current `main`: `datafusion.user_defined.WindowEvaluator` already provides the capability. `crates/core/src/udwf.rs::RustPartitionEvaluator` dispatches `memoize` / `get_range` / `evaluate` / `evaluate_all` / `evaluate_all_with_rank` / `supports_bounded_execution` / `uses_window_frame` / `include_rank` via `call_method`, and `python/tests/test_udwf.py` has several pure-Python subclasses driven through `udwf()`.

It's deliberately not `metaclass=ABCMeta` like `Accumulator`, though: which `evaluate*` method must be implemented depends on the `uses_window_frame` / `supports_bounded_execution` / `include_rank` flag matrix documented on the class, so no single method can be marked `@abstractmethod` the way `Accumulator`'s can.

Opening this to confirm: is the literal "ABC" form still wanted on top of what exists (and if so, what should the abstractness contract look like given the flag matrix), or should #1577 drop this item?

閱讀 datafusion.user_defined.WindowEvaluator 和 crates/core/src/udwf.rs::RustPartitionEvaluator，然後檢查 python/tests/test_udwf.py 中的純 Python 子類別。確定是否仍然需要 ABC；如果需要，則定義一個涵蓋已記錄 flag 矩陣的抽象性契約；當所要求的方向得到解決並反映在 issue 或實作計畫中時，即視為完成。

由索引模型根據 Issue 內容生成。