Two related multi-band accuracy bugs from the geotiff audit:

A2 (HIGH): GPU tiled reads of planar=2 files (separate band planes)
returned wrong pixel values. gpu_decode_tiles iterates a single
TileOffsets list with bytes_per_pixel = itemsize * samples and assumes
chunky interleave. For planar=2, each band has its own contiguous run
of tiles in TileOffsets, so the kernel read across band boundaries and
produced garbage. read_geotiff_gpu now detects planar==2 and decodes
each band's tile slab as a single-band image, then stacks the results
into (H, W, samples). The chunky planar=1 path is unchanged.

A3 (MEDIUM): The GPU stripped multi-band fallback hardcoded
dims=['y','x'] even for 3-D CPU reads, mirroring the pre-xarray-contrib#1525 bug.
The same fix lands here so the planar=2 stripped tests pass; this
matches the change in PR xarray-contrib#1525 verbatim.

Both paths assert the assembled shape equals (H, W, samples) before
returning, so a future kernel regression surfaces immediately rather
than producing garbage.

Adds xrspatial/geotiff/tests/test_planar_multiband.py covering 53
combinations: planar in {separate, contig} x layout in {tiled, stripped}
x bands in {2, 3, 4} x dtype in {uint8, uint16} x backend in
{CPU, GPU}, plus single-band sanity checks and an explicit A3 repro.
GPU tests skip cleanly when CUDA is unavailable. Seeds are fixed.

Refs audit issues A2, A3.

* Handle PlanarConfiguration=2 across CPU + GPU multi-band reads

Two related multi-band accuracy bugs from the geotiff audit:

A2 (HIGH): GPU tiled reads of planar=2 files (separate band planes)
returned wrong pixel values. gpu_decode_tiles iterates a single
TileOffsets list with bytes_per_pixel = itemsize * samples and assumes
chunky interleave. For planar=2, each band has its own contiguous run
of tiles in TileOffsets, so the kernel read across band boundaries and
produced garbage. read_geotiff_gpu now detects planar==2 and decodes
each band's tile slab as a single-band image, then stacks the results
into (H, W, samples). The chunky planar=1 path is unchanged.

A3 (MEDIUM): The GPU stripped multi-band fallback hardcoded
dims=['y','x'] even for 3-D CPU reads, mirroring the pre-#1525 bug.
The same fix lands here so the planar=2 stripped tests pass; this
matches the change in PR #1525 verbatim.

Both paths assert the assembled shape equals (H, W, samples) before
returning, so a future kernel regression surfaces immediately rather
than producing garbage.

Adds xrspatial/geotiff/tests/test_planar_multiband.py covering 53
combinations: planar in {separate, contig} x layout in {tiled, stripped}
x bands in {2, 3, 4} x dtype in {uint8, uint16} x backend in
{CPU, GPU}, plus single-band sanity checks and an explicit A3 repro.
GPU tests skip cleanly when CUDA is unavailable. Seeds are fixed.

Refs audit issues A2, A3.

* Address Copilot review on #1532

Seven findings, all acted on:

- Sparse-tile bypass: planar=2 branch ran before the has_sparse_tile
  check, so a planar=2 file with sparse tiles would hit the GPU path
  that doesn't handle sparse blocks. Add `not has_sparse_tile` to the
  planar=2 gate so sparse always routes to the CPU reader.
- Strict offset-count check: planar=2 required len(offsets) ==
  tiles_per_band * samples, but validate_tile_layout only requires
  >=. Files with extra trailing entries that pass the CPU reader were
  rejected here. Relax to a minimum-length check and slice the first
  tiles_per_band * samples entries.
- Auto-mode fallback semantics: stage-2 GPU failure in
  _gpu_decode_single_band_tiles raised unconditionally instead of
  letting the caller fall back to a whole-image CPU read like the
  chunky path does. Helper now returns None on stage-2 auto failure;
  caller catches the signal, runs read_to_array + cupy.asarray, and
  matches the chunky path's gpu='auto' contract. Strict mode still
  re-raises.
- Per-band file re-reads: the helper called src.read_all() for each
  band's stage-2 fallback, doing N redundant full-file reads on a
  planar=2 image with N bands. Caller now does read_all() once before
  the band loop and passes the bytes via a new shared_data parameter.
- Unused params: drop overview_level and cupy from the helper
  signature; neither was referenced.
- Two assert statements for runtime shape validation replaced with
  explicit `if ...: raise RuntimeError(...)` so the checks survive
  python -O.

53 planar tests + 303 multi/band/planar/strip/tile/sparse regression
tests still pass.

* Switch shared file read to lazy reader

The previous fix did read_all() unconditionally before the band loop
to amortise a stage-2 fallback across bands. When every band's GDS
path succeeds, those bytes are never used. Replace the eager
read_all() with a closure that materialises the buffer on first call
and caches it, so:

- Every-band-GDS-succeeds case: 0 read_all() calls (was 1).
- Any-band-falls-back case: exactly 1 read_all() call regardless of
  how many bands fall back (was N).

Same Copilot-finding-#4 fix, just lazier.