Use public Sentinel-2 COGs

[snamber]

Summary

• discover Sentinel-2 L2A scenes through Tilebox's open_data.aws_earth.sentinel2 dataset
• read semantic assets from the public AWS COG archive with Tilebox storage, removing Copernicus credentials and manual JP2 traversal
• update the Tilebox Python packages to 0.62.0 and use async workflow execution for concurrent band reads
• preserve the projected 3 km crop footprint when deriving each raster window

Validation

• uv lock --check
• uv run ruff check .
• uv run ruff format --check .
• uv run ty check
• uv run python -m compileall -q datacenters
• WORKFLOW_CACHE_BUCKET='' tilebox workflow build-release --json
• real AWS COG read: Creekstone scene, 301×301 pixels, 0% invalid data, 0% crop cloud
• 32 GiB end-to-end smoke job 01a0c448-6a3d-df9e-4cc4-0b3c725f8a44: 5/5 tasks computed, ranking.json produced, Creekstone ranked first with score 90.8151