Align design-point inputs with model runs and make site selection reproducible

[dlebauer]

Follow-up to #19.

PR 19 generates 1,000 design points from the consolidated parcels. That is sufficient for work to proceed.

The following are lower priority follow on steps:

1. Use consistent source inputs for site selection and model runs. In particular, the PR uses tillage v1.0 while v2.0 is available. This can be addressed alongside "use existing config framework" task below.

2. Separate design points from anchor sites. Implement separate design points and anchor sites organization#263: select the design independently, retain a separate anchor list, and run the deduplicated union with both roles recorded. Use design points for downscaling, use anchors for evaluation. This will also be necessary (later) to ensure that the design point number vs model skill analysis is interpretable Analyze design point number vs downscaling model skill organization#192.

3. Use the downscaling cli config for clustering + site selection. Clustering + site selection does not require CLI integration, but this can help ensure 1) all files are available, and 2) the same files are used in the design point selection and downscaling workflows. Retire 000-config.R.

4. Put inputs on S3. Inventory the ERA5, TWI, SoilGrids, county, and climate-region inputs used by script 010. Stage the exact required products in versioned S3 locations instead of hard-coded geo paths.

5. Fix the crop-standardization tests. Update the fixture in tests/testthat/test-standardize_cadwr_crops.R to use the LandIQ lookup schema (legend_year, PFT, CLASS, SUBCLASS, SUBCLASS_desc). All five existing tests currently error because the fixture still uses the previous schema. Rerun the tests and add focused coverage for subclass normalization, conflicting PFT mappings, and the cropped/uncropped selection threshold.