The reader was ignoring tag 274 and returning the file's stored pixel
order regardless of which corner the data was meant to start from. Files
written with orientation 2-8 came back as silently flipped or rotated
arrays.

Parse the tag in `IFD.orientation` (default 1 = top-left, no transform)
and apply the TIFF 6.0 spec table after decode in `_apply_orientation`.
Orientations 5-8 swap rows and columns, so the GeoTransform's pixel_width
and pixel_height swap too -- coords end up with the right size on each
axis after the remap.

A windowed read on a non-default orientation has ambiguous semantics
(does the window refer to file pixels or display pixels?), so the
combination raises ValueError rather than silently picking one.

Tests cover all eight orientations against a manually-computed expected
output, the dim-swap for 5-8, default-tag behaviour, and the window
combination raising. tifffile's `imread` does not itself apply the tag,
so the comparison target is computed from the source array; that's the
only reliable way to verify the spec without depending on a second
oriented decoder.

… band order, error msg

Six review findings on the orientation fix; all addressed in one commit.

- ``TAG_ORIENTATION`` now in ``_MANAGED_TAGS`` so it's not collected into
  ``extra_tags`` on read. Previously, reading an orientation=4 file then
  writing it back through ``to_geotiff`` re-emitted the original tag and
  the next read applied the orientation a second time -- doubly-flipped
  output. Round-trip now stable.
- The transform swap for orientations 5-8 used
  ``pixel_width=abs(t.pixel_height), pixel_height=-abs(t.pixel_width)``
  which silently coerced any input to a north-up convention. South-up or
  west-up files would flip direction. Now does a plain swap that
  preserves the original signs. For orientations 6/7/8 on georeferenced
  files the swap is shape-correct but geometrically approximate -- a
  warning is now emitted so the user knows to verify the GeoTransform.
- Single-band slicing (``band=N``) now happens before
  ``_apply_orientation`` so reorientation runs on a 2D array rather
  than a multi-band cube that gets sliced afterwards.
- Dropped the redundant local ``from ._geotags import GeoTransform``
  inside the read path; the symbol is already imported at module scope.
- Improved the ValueError for ``window=`` + non-default orientation:
  includes the actual orientation value and notes that dask-chunked
  reads (``chunks=...``) hit the same path.
- New tests cover: tag 274 not surviving to extra_tags for any of
  orientations 2-8, a round-trip read+write+read returning the same
  array (regression for the doubled-application bug), the warning
  firing on georeferenced 5-8 files, and band selection + orientation
  returning the expected 2D array.