geotiff: align direct backend mask_nodata default with open_geotiff

[brendancol]

Closes #2976

What

The three direct GeoTIFF backend entry points defaulted to mask_nodata=True while the public open_geotiff defaults to masked=False. A bare backend call returned a different dtype and NaN-substituted values than the public path, silently.

• Flip _read_geotiff_dask, _read_geotiff_gpu, and _read_vrt to mask_nodata=False so a direct call with no mask_nodata argument matches open_geotiff (unmasked, rioxarray-compatible).
• The public wrapper and the internal GPU-to-dask / dask-to-VRT callers all forward mask_nodata explicitly, so only direct callers are affected.

Backend coverage

The dask and VRT entry points cover numpy and cupy graphs; the GPU entry point covers eager and dask+cupy. The default flip applies the same way on all of them.

Test plan

• New regression tests: each direct backend with no mask_nodata produces the same unmasked result (dtype + sentinel preserved, no NaN) as open_geotiff. The GPU test is guarded by requires_gpu.
• Updated existing tests that relied on the old masked default. Where a test verifies the masking path it now passes mask_nodata=True; where a test pinned the bare-call behavior it asserts the new unmasked default.
• Full geotiff suite green locally (6155 passed, GPU included).
• flake8 clean on touched files.

[brendancol]

PR Review: geotiff: align direct backend mask_nodata default with open_geotiff

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

One note, not blocking. The internal helpers that aren't public entry points kept mask_nodata: bool = True: _read_geotiff_gpu_chunked_gds in gpu.py, plus _vrt_chunk_read and the chunked dispatch in vrt.py. That's correct. They always get an explicit mask_nodata forwarded from the entry point and aren't a caller's direct surface. The tests in test_reader.py and test_kernels_and_kwargs.py now pass mask_nodata=True to both sides when they compare an entry point against an internal helper, so those comparisons stay apples-to-apples. Flagging it only so the next reader doesn't mistake the asymmetry for an oversight.

Nits (optional improvements)

None.

What looks good

• The source change is small: three default flips and the matching default True to default False docstring edits. Everything else in the diff is test updates.
• The public wrapper forwards mask_nodata=masked on every dispatch path (GPU, dask, VRT, eager), and the two internal cross-backend callers (gpu.py:1422 to _read_geotiff_dask, dask.py:227 to _read_vrt) forward it explicitly. So the default only affects direct callers. Confirmed by reading the call sites.
• The new regression tests check direct-backend-vs-open_geotiff parity for all three backends (GPU guarded by requires_gpu): dtype, sentinel survival, no NaN, bit-equal pixels.
• Existing tests that pinned the old masked default were split the right way. Masking-path tests now request mask_nodata=True; default-behavior tests assert the new unmasked default and were renamed for #2976.
• Full geotiff suite passes locally including GPU (6155 passed).

Checklist

• Behavior change matches the issue: direct backend default now matches open_geotiff
• All three backends produce consistent unmasked results with no mask_nodata arg
• NaN / sentinel handling correct: masking still available via mask_nodata=True
• Edge cases covered: int sentinel hit, float sentinel, dtype preservation, dask + GPU graphs
• Internal helper defaults left untouched (correct: they receive explicit values)
• No premature materialization or new copies (no compute path changed)
• Benchmark not needed (no perf-critical path changed)
• README feature matrix not applicable (no new public function)
• Docstrings updated to default False on all three backends