fix: correct WMA calculation to use nansum instead of nanmean

[Ayush10]

Summary

Closes #1993

The weighted_mean function in WMA._load_internal (qlib/data/ops.py) normalizes weights to sum to 1 via w = w / w.sum(), then incorrectly uses np.nanmean(w * x) which divides by the element count again. The correct function is np.nansum(w * x) since the weights already sum to 1.

This is consistent with how EMA.exp_weighted_mean is implemented (which already uses np.nansum).

Before

def weighted_mean(x):
    w = np.arange(len(x)) + 1
    w = w / w.sum()
    return np.nanmean(w * x)  # BUG: divides by count again

After

def weighted_mean(x):
    w = np.arange(len(x)) + 1
    w = w / w.sum()
    return np.nansum(w * x)  # correct: weights already sum to 1

Test Plan

• Change is consistent with EMA implementation pattern
• Single-line fix, minimal risk