[SeriesDFilter] Redundant execution of filter_pipe for each fields_group causes slow data loading

[CanvaChen]

Description

During data loading, I noticed that filter_pipe is executed repeatedly for each fields_group instead of being applied once or cached. This causes significant slowdowns in data preparation, especially when multiple field groups are configured.

Steps to Reproduce / Observation

1. Added logging/print statements inside the _getFilterSeries method.
2. Ran the data loading pipeline with multiple fields_groups (e.g., 3 groups in my setup).
3. Observed that the filter pipeline was called exactly 3 times, once per fields_group, rather than sharing the filtered result across groups.

[he-yufeng]

I found a small loader-side fix for this: resolve ilter_pipe once at the QlibDataLoader.load() level, then reuse the filtered instruments across field groups. I added a regression test that checks D.instruments(..., filter_pipe=...) is called once while grouped feature loading still runs per group. I'll open the PR shortly.

[CanvaChen]

是因为第一个提交仍然会多次执行过滤，所以做了第二次提交吗？

[CHGY003]

是的,第一次提交仍然会在某些情况下重复执行过滤,所以我在第二次提交中进一步修复。

[CanvaChen]

使用 task.dataset.kwargs.handler 如 Alpha158，仍然过滤多次，日志如下：

[334838:MainThread](2026-06-04 21:31:42,951) INFO - qlib.data - [data.py:100] - 2026-01-05 00:00:00 ~ 2026-03-31 00:00:00 5439 过滤比例：29444/304584
[334838:MainThread](2026-06-04 21:33:05,538) INFO - qlib.data - [data.py:100] - 2026-01-05 00:00:00 ~ 2026-03-31 00:00:00 5439 过滤比例：29444/304584
[334838:MainThread](2026-06-04 21:34:18,894) INFO - qlib.timer - [log.py:127] - Time cost: 158.973s | Loading data Done

[he-yufeng]

Thanks, this pointed out the missing layer in the PR.

I pushed another update to #2243. The first version only reused the stockpool config from D.instruments(...), but each grouped D.features(...) call still resolved that config through list_instruments(...), so feature and label could still run the dynamic filter separately.

The latest commit resolves a stockpool config with filter_pipe to the filtered instrument dict once per grouped loader/frequency, then reuses that dict across field groups. The Alpha158 regression now checks D.list_instruments(...) is called once, not just D.instruments(...).