Record accuracy sweep results for mahalanobis

[brendancol]

Summary

• Audited xrspatial/mahalanobis.py against the five accuracy categories (precision loss, NaN/Inf, off-by-one, projection, backend parity).
• No critical or high issues. All four backends cast to float64, filter NaN/Inf via isfinite, clamp dist_sq non-negative before sqrt, and raise ValueError on exactly singular covariance.
• Recorded one LOW note: np.linalg.inv has no pseudo-inverse fallback so a near-singular (but not exactly singular) covariance can return garbage without raising; the two-pass mean/cov can also lose precision for inputs with very large mean and small variance. Both are documented but not fixed.

Test plan

• csv updated via csv.DictReader/Writer
• no source files touched