geotiff: add attrs['masked_nodata'] alongside attrs['nodata'] (#1988)

[brendancol]

Summary

Closes the foundational chunk of #1988: split the declared file sentinel from the masked-state signal in DataArray attrs on every geotiff read path.

• attrs['nodata'] -- declared file sentinel as a scalar of the source dtype. Set whenever the source declared one. Independent of whether masking ran.
• attrs['masked_nodata'] -- new boolean. True iff the reader replaced sentinel pixels with NaN in the in-memory array (final dtype is float). False iff the array still carries the literal integer sentinel.

The two signals were previously fused: attrs['nodata'] doubled as "the file declared this sentinel" and "the reader already replaced sentinel pixels with NaN." Downstream code branching on the second meaning had to use the first as a proxy and lost the ability to tell a float-NaN array from an int-with-sentinel array.

Changes

• New _set_nodata_attrs(attrs, nodata, *, array_dtype) in _attrs.py. Sets both keys atomically from the final array dtype.
• Every read backend calls it from one site, after any masking and any user-requested dtype cast:
  • open_geotiff eager numpy (__init__.py)
  • read_geotiff_dask (_backends/dask.py) -- uses the declared graph dtype
  • read_geotiff_gpu -- three sites (eager tiled, eager stripped, chunked dask+gpu) in _backends/gpu.py
  • read_vrt eager + chunked in _backends/vrt.py
• Test matrix in test_nodata_semantics_split_1988.py: source-has-sentinel x backend x {float-NaN-masked, int-sentinel-preserved}, plus direct unit coverage of _set_nodata_attrs.

What this does not change

This PR is additive. The writer still reads attrs['nodata'] through _resolve_nodata_attr and ignores masked_nodata; no existing nodata test changes its expectation. Migrating writers and downstream xrspatial code to branch on masked_nodata is the follow-up scope on #1988.

Test plan

• pytest xrspatial/geotiff/tests/test_nodata_semantics_split_1988.py (12 passed)
• All existing nodata regression suites pass: test_nodata_nan_int_1774, test_nodata_int64_precision_1847, test_nodata_attr_aliases_1582, test_miniswhite_nodata_1809, test_vrt_band_nodata_1598, test_overview_nodata_inheritance_1739, test_vrt_int_source_float_dtype_1616, test_vrt_multiband_int_nodata_1611 (88 passed)
• Full xrspatial/geotiff/ suite (excluding the pre-existing fuzz failure on test_fuzz_hypothesis_1661.py, which reproduces on main): 952 passed

[brendancol]

Self-review

Blockers

None.

Suggestions addressed in follow-up commit f7e6cb6

• xrspatial/geotiff/tests/test_nodata_semantics_split_1988.py:29 had an unused import xarray as xr. Dropped.
• The original test file had no GPU or VRT coverage despite the PR touching five sites across _backends/gpu.py (three) and _backends/vrt.py (two). Added:
  • TestVRTEager (3 tests): float-typed VRT over int source with hit, int-typed VRT with no hit, VRT band with no <NoDataValue>.
  • TestVRTChunked (1 test): declared-float64 graph for in-range int sentinel.
  • TestGPU (3 tests, guarded by _gpu_only): read_geotiff_gpu on int source with hit, int source no-hit (stays int), and the dask + GPU path with in-range sentinel.

Nits (left as-is)

• xrspatial/geotiff/tests/test_nodata_semantics_split_1988.py:189 -- the hand-crafted uint16 TIFF builder duplicates _build_uint16_tiff from test_nodata_nan_int_1774.py:54. Local self-documenting duplication, leave for a future sweep.
• xrspatial/geotiff/_attrs.py:108 -- bool(np.dtype(array_dtype).kind == 'f') -- the bool(...) wrapper is technically redundant (== on strings already yields a Python bool) but harmless and self-documenting.

What looks good

• Single helper _set_nodata_attrs in _attrs.py keeps the contract in one place; every read site delegates to it.
• The masked_nodata signal is derived from the final array dtype, which is the right invariant to anchor on.
• Cross-backend divergence is intentional and documented: eager numpy stays int when the sentinel does not hit (masked_nodata=False), while dask declares float64 up front for any in-range sentinel (masked_nodata=True). Both reflect their actual array dtype, so the contract masked_nodata == (dtype is float) holds end-to-end.
• The change is additive. No existing nodata test had to be modified. The writer side is untouched and the next geotiff: split declared nodata from masked-data state in DataArray attrs #1988 follow-up can migrate _resolve_nodata_attr / to_geotiff to consult masked_nodata.

Final test counts

• New file: 19 tests, all passing locally (incl. 3 GPU tests on a cupy+CUDA env).
• Existing nodata regression suites: 88 passing.
• Full xrspatial/geotiff/ suite (excluding pre-existing test_fuzz_hypothesis_1661.py failure that reproduces on main): 952 passing.