contours: keep CRS on empty geopandas result (#2700)

[brendancol]

Closes #2700

contours(agg, return_type="geopandas") crashed when the input raster had a crs attribute but the requested levels produced no contour lines (flat raster, levels outside the data range, etc.). _to_geopandas() called gpd.GeoDataFrame(records, crs=crs) with an empty records list, and geopandas rejects a CRS on a frame that has no geometry column.

Changes:

• _to_geopandas() builds an empty frame with an explicit geometry column when there are no records, so the input CRS still attaches and the result is a well-formed (empty) GeoDataFrame.
• The all-NaN early-return path now passes the input CRS instead of None, so it no longer silently drops georeferencing.

Backend coverage: the geopandas conversion runs on the host after tracing, so this path is identical for numpy, cupy, dask+numpy, and dask+cupy. No backend-specific code changed.

Test plan:

• New: empty result with a CRS returns an empty GeoDataFrame carrying the CRS (no crash)
• New: all-NaN input with auto levels keeps the CRS
• New: populated result still carries the CRS
• New: empty result with no input CRS returns an empty frame without error
• pytest xrspatial/tests/test_contour.py -> 28 passed

[brendancol]

PR Review: contours: keep CRS on empty geopandas result (#2700)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/contour.py:551 - crs is passed twice in the empty branch: once to gpd.GeoSeries([], crs=crs) and once to the outer GeoDataFrame(..., crs=crs). The outer one sets the frame CRS, so the inner one is redundant. It is harmless and guarantees the geometry column itself carries the CRS, so dropping it is optional.
• The empty-frame level column comes from [], so its dtype is object/float rather than matching the populated path. This only matters if a caller concatenates an empty result with a populated one and checks dtypes. Low impact.

What looks good

• The fix targets the exact failure: an empty records list plus a non-None CRS. Building the frame with an explicit geometry column is the right way to let geopandas attach the CRS.
• The all-NaN early-return now forwards the input CRS instead of None, which closes the silent-drop path.
• Four new tests cover the populated, empty-with-CRS, all-NaN, and empty-without-CRS cases. The empty-with-CRS test names the regression (#2700).
• The change is surgical, with no unrelated refactoring.

Checklist

• Algorithm matches reference: n/a (CRS propagation fix)
• All implemented backends consistent: yes, conversion is host-side after tracing
• NaN handling correct: yes, all-NaN path verified by test
• Edge cases covered by tests: yes
• Dask chunk boundaries: n/a
• No premature materialization: yes
• Benchmark: not needed
• README feature matrix: not needed
• Docstrings present and accurate: yes, no API change

[brendancol]

Follow-up review (after 5d46d52)

Addressed the first nit: dropped the redundant crs=crs on the inner gpd.GeoSeries([]). The outer GeoDataFrame(..., crs=crs) still sets the frame CRS, and test_geopandas_empty_result_keeps_crs confirms the empty frame carries it. All 28 contour tests pass.

The second nit (empty-frame level column dtype) is left as is. Forcing a dtype on a zero-row column adds noise for an edge case that only matters under cross-frame concatenation with dtype checks, which the library does not do here. Recording it rather than fixing.

No remaining blockers or suggestions.