extracting surface heights from optical images requires the availability of images acquired from different angles. I don’t think the spatial resolution of Sentinel-2 images allows to exploit, or even to notice, the required parallax. According to the documentation, the shifts occur at a sub-pixel level and are corrected in the pre-processing of Sentinel-2 products already.
There are studies which use it for the detection of moving objects or clouds (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5992602/), but I am not aware of any possibility to retrieve reliable elevation data, at least not with the same methods which are applied to IRS or Pléiades (classic photogrammetry or structure from motion).
Please someone correct me if I miss a point.