🏁 Correct shift of training tiles by half a pixel

Adjusting the bounding box positions of (some) tiles by half a pixel (roughly 500m). Fixed a poorly written unit test for the data_prep.get_window_bounds function introduced in ab0295e right before release v0.4.0, and refactored the function to make it work correctly with newer libraries. Strangely enough, after recreating the geojson tiles, the coordinate changes are only for the Operation Ice Bridge/CRESIS Basler and TO grids, and we still have exactly 2347 tiles!

Specifically, we've switched from using xarray.open_rasterio to xarray.open_dataarray as xarray's rasterio backend fails to read NetCDF georeferencing information properly with rasterio >= 1.0.14 (see #84). The projection information is monkey-patched into the xarray grid using salem. Most important bit is to ensure that we go from xarray's centre-based pixel coordinate system to rasterio's corner-based pixel coordinate system (accomplished with dataset.salem.grid.corner_grid) via a manually set Affine transformation.