fix: mask degenerate expanding rsquare windows

[Parafee41 (koriyoshi2041)]

Problem

Rsquare(feature, 0) uses the expanding Cython kernel but, unlike the rolling path, does not mask windows where the feature standard deviation is effectively zero. Floating-point cancellation can therefore leak inf or misleading finite R-squared values into expanding features.

Fixes #2297.

Fix

Apply the same std ~= 0 guard used by rolling R-squared to the expanding path, using Series.expanding(min_periods=1).std() so the mask follows each expanding window.

Test

• Added a regression for the reported near-constant series; every degenerate expanding window is now NaN.
• Added a varying-series case to confirm valid expanding R-squared values remain finite.
• python -m pytest tests/ops/test_rolling_operator.py -q — 2 passed.
• Black check on the new test file and git diff --check passed.

Risk

Low. The change is limited to the N == 0 R-squared branch and reuses the existing rolling-path tolerance (atol=2e-05). Valid non-degenerate windows and all other rolling operators are unchanged.

[Parafee41 (koriyoshi2041)]

Closing as duplicate. #2298 already contains the same fix plus regression coverage, and #2310 also covers the code change. I should have caught those existing PRs before opening this one.