Change rounding method for low-resolution / smooth_frames?

The method used by WFDB to downsample signals in "low-resolution mode" is inaccurate (samples are rounded towards zero rather than to the nearest integer.) I'm intending to improve this by using a more accurate method in WFDB 10.7:

https://github.com/bemoody/wfdb/commit/bbfecc133932ae61f3c32491fe4679067b75c9ec

The method used by WFDB 10.6 and earlier is (as far as I know) equivalent to what rdrecord currently does by default (i.e. with smooth_frames=True).

I'm looking for opinions here. Should we make the same change in wfdb-python for the sake of better accuracy? Should we keep the existing inaccurate rounding for the sake of backward reproducibility?

(Implementing this change efficiently with numpy might be a little difficult, but let's ignore that for now.)

Beginne mit dem rdrecord-Pfad, der bei smooth_frames=True verwendet wird, und vergleiche dessen Rundungsverhalten mit dem verlinkten WFDB-10.7-Commit. Kläre vor jeder Änderung, ob Genauigkeit oder Rückwärtsreproduzierbarkeit Vorrang hat, und evaluiere anschließend einen effizienten NumPy-Ansatz und definiere die erwarteten Ergebnisse bei niedriger Auflösung.