feat(download): add optional GDAL-native labels export backend

[JesJehle]

Problem

Large prediction exports (e.g. dataset 665 with ~374k geometries) are currently generated via Python feature materialization and GeoPandas/Fiona writes. This can take a long time and may remain in processing for extended periods.

Goal

Add an optional GDAL-native export backend for labels downloads to improve throughput and stability for large datasets, while keeping the current Python exporter as a fallback.

Proposed approach

• Add a configurable exporter backend flag (e.g. LABEL_EXPORT_BACKEND=python|gdal), defaulting to python initially.
• Implement a GDAL-based exporter path in the API background task using ogr2ogr.
• Build deterministic SQL layer exports for a dataset into one GPKG:
  • deadwood_model_prediction
  • deadwood_visual_interpretation
  • forest_cover_model_prediction
  • forest_cover_visual_interpretation
  • aoi
• Keep existing route/error contract (.error marker file, status endpoint returns failed with message).
• Add timeout + robust subprocess logging (stdout/stderr + elapsed time).

Configuration

• Do not read DSN from local MCP config files.
• Introduce explicit environment variable for export DSN (e.g. GDAL_EXPORT_PG_DSN) managed via deployment secrets.
• Prefer direct local Postgres endpoint if available for large export workloads.

Testing plan

• Unit tests:
  • command/SQL builder for each layer
  • dataset/source filtering correctness
• Integration tests (api-test container):
  • export completes
  • expected layers exist in output GPKG
  • feature counts match DB counts for fixture dataset
• Failure-path tests:
  • invalid DSN / timeout -> .error marker + failed status response
• Performance smoke benchmark:
  • compare current Python path vs GDAL path on a large dataset and record wall-clock time.

Acceptance criteria

• New backend can be enabled with config toggle.
• Endpoints remain backward compatible.
• Export succeeds with correct layer schema/content.
• Clear error propagation on failure.
• Documented rollout plan to switch default backend after validation.

Out of scope

• Pre-generating/caching exports.
• Changing download response format away from GPKG.

[linear]

DT-278