fix: compute WMA as weighted sum

[he-yufeng]

Summary

Fixes #1993.

WMA already normalizes the rolling weights so they sum to 1, but then calls np.nanmean(w * x). That divides the weighted values by the number of non-NaN entries again. For example, [1, 2, 3] with a 3-day window should be (1 + 4 + 9) / 6 = 14/6, while the current implementation returns 14/18.

This changes the reducer to np.nansum(w * x), matching the normalized-weight definition and the behavior used by the exponential weighted reducer nearby.

Validation

python setup.py build_ext --inplace

$env:PYTHONPATH=(Resolve-Path .).Path; python -m pytest tests\ops\test_rolling_ops.py -q
# 1 passed

$env:PYTHONPATH=(Resolve-Path .).Path; python -m pytest tests\ops\test_special_ops.py -q
# 2 passed

python -m py_compile qlib\data\ops.py tests\ops\test_rolling_ops.py
python -m ruff check qlib\data\ops.py tests\ops\test_rolling_ops.py
# All checks passed

git diff --check