Extract the classification schemes out of sampleearth.landcover, where they
lived as Python tables, into commented TOML that a domain expert can review in
a diff.

- data/  two schemes: se-landcover-list (70 classes) and se-cover-facets
  (82 components, 18 tag axes), each a directory of TOML
- models.py  dataclasses; the hierarchy, crosswalk and LCCS dichotomous
  derivation are methods on the scheme, not functions over a global table, so
  two schemes can be loaded at once
- loader.py  tomllib + every validator the Python tables used to enforce.
  Rules stay in code: moving the tables out was about reviewable diffs, not
  about letting a typo reach a GeoPackage
- export.py  JSON bundle, JSON Schema and per-table CSV. The bundle carries
  the LCCS decision tree generated from the same data, so a rendering of it
  cannot say something derive_dichotomous would not

No runtime dependencies: tomllib is stdlib from 3.11, which is what lets the
QGIS plugin vendor this the way it vendors sampleearth_core.

92 tests, including a parity suite proving the extraction lost nothing against
the sampleearth tables it came from. That suite is deleted once sampleearth
consumes this package instead of its own copy.