Round-trip transform, crs, and tag metadata through to_geotiff/open_geotiff

[brendancol]

Polish GeoTIFF metadata so transforms, CRS, and tag metadata round-trip cleanly through to_geotiff / open_geotiff. Findings M-1 through M-4 from the geotiff metadata audit.

M-1 / M-2: transform and CRS type stability

open_geotiff() stores attrs['crs'] as an int EPSG code. read_vrt() does the same when the EPSG can be derived from the WKT, otherwise it leaves the WKT in attrs['crs_wkt']. The two paths agree, but neither docstring says so. Document the convention in both functions and in _geo_to_coords. Keep tolerating string CRS values on write so existing callers don't break.

For the transform: open_geotiff() builds y/x coords from the file's GeoTransform, and to_geotiff() reconstructs the transform from coords via _coords_to_transform. That round-trips fine for clean numbers but drifts on fractional or precision-edge transforms. Stash the rasterio 6-tuple on read as attrs['transform'] so to_geotiff() can reuse it directly, falling back to _coords_to_transform only when the attr is missing.

M-3: ColorMap, ExtraSamples, ImageDescription pass-through

_geotags.py already reads ColorMap (exposed as attrs['cmap'] / attrs['colormap_rgba']) but ignores ExtraSamples and ImageDescription. Capture all three on read into attrs['colormap'] (raw uint16 triple), attrs['extra_samples'], and attrs['image_description']. On write, route them through the existing extra_tags pass-through when present.

M-4: integer-with-nodata dtype promotion

open_geotiff() silently promotes integer rasters with a nodata sentinel to float64 with NaN replacing the sentinel. This is documented nowhere. Add it to the docstring along with the dtype=... override.

Tests

• Round-trip attrs['crs'] and attrs['transform'] through write -> read -> write -> read.
• ColorMap round-trip on a uint8 indexed raster.
• ImageDescription round-trip with a known string.
• uint16 with nodata=65535 reads back as float64 with NaN at the sentinel; passing dtype='uint16' raises ValueError (existing float-to-int guard).

Same polish that landed in reproject (#1462) and resample (#1472).