feat: Add test suite, unit tests on PR, and weekly live-service tests

[JoshCederstrom]

This pull request adds a test suite and continuous-integration coverage to pytRIBS. Until now the GitHub Actions workflow only installed the package and imported it. Multiple failure modes are now covered by tests, in two tiers: fast unit tests that run on every pull request, and scheduled live-service tests that run weekly against the real external APIs.

Unit tests (run on every PR)

• test_rosetta_soil_params.py: direct regression guard. Feeds four textbook textures (sand, loam, clay, silt loam) through _rosetta_to_tribs_params and asserts (1) literature-informed plausibility windows on Ks per texture, physical bounds on theta_r/theta_s/psib/m, and correct texture ordering and (2) "golden values" frozen from rosetta-soil 0.3.x at a loose ±25% tolerance to flag subtler prediction drift. Also verifies the ROSETTA model-code selection.
• test_soil_map.py: USDA texture classification through create_soil_map, covering both the ISRIC (g/kg) and SOLUS (% mass) input conventions and the sequential class renumbering.
• test_ks_decay.py: generates synthetic Ks-with-depth grids from a known decay parameter and asserts compute_ks_decay recovers it, floors uniform profiles at min_f, and propagates nodata.
• test_met_conversions.py: hand-checked "golden values" for the NLDAS unit conversions (Pa to hPa, K to degC, specific humidity to RH, 10 m to 2 m wind log-profile scaling), the GMT timestamp shift, .mdf/.sdf round trips, and the observation-station validation errors.
• test_soil_tables.py / test_ascii_io.py: round trips for the .sdt soil table and the ESRI ASCII raster writer/reader, including NaN-to-nodata replacement and cell-size preservation.
• test_evaluate_metrics.py: known-answer checks for NSE, KGE, RMSE, and percent bias.
• test_classes.py: constructs every top-level class bare and from a written .in file, locking in the constructor fix and the input-file write/read round trip.

Live-service canaries (automatic, weekly)

Small real requests with plausibility gates on the responses, so an upstream API change is flagged by a failed-workflow email instead of by a broken user run:

• ISRIC SoilGrids WCS, SOLUS100 COG windowed reads, and POLARIS tile downloads over a ~2×2 km test domain, each asserting a readable, non-empty grid with in-range values.
• NASA Giovanni NLDAS point timeseries, asserting all seven forcing variables return with physically plausible values. Because get_nldas_point swallows per-variable fetch errors, a missing column is the failure signal.

Continuous integration

• install_and_test.yml: now installs .[test] and runs pytest (network tests excluded by marker) after the import smoke test, on Python 3.11 and 3.13. Because dependencies are unpinned, every PR run also exercises the newest dependency releases.
• canary.yml (new): scheduled Mondays 12:00 UTC plus manual dispatch. Installs the latest release of every dependency and runs the full unit suite, so a breaking dependency release is caught within a week even with no pushes. The Giovanni test authenticates via the EARTHDATA_USERNAME/EARTHDATA_PASSWORD repository secrets.

Technical changes

• tests/ (new): nine test modules plus a conftest.py fixture that writes small georeferenced ESRI ASCII grids for the raster-based tests.
• pytRIBS/classes.py: fixed the broken input_file path in the Soil, Land, Met, and Mesh constructors described above; each now also carries a populated options dictionary, consistent with Results.
• pyproject.toml: added a test extra (pytest>=8.0) and pytest configuration.
• .github/workflows/install_and_test.yml: unit test step added.
• .github/workflows/canary.yml: new scheduled canary workflow.

Running the tests

pip install -e ".[test]"

pytest              # fast unit suite (network canaries excluded)
pytest -m network   # live-service canaries (Giovanni needs Earthdata credentials)