Replace 1.20.6 with 1.20.5; fix broken common() left over from the Pooch

refactor

Why

Two latent issues were hiding behind soft-skipped tests:

1. 1.20.6 is declared as a supported version but upstream ships no data for
   it. The PrismarineJS minecraft-data tarball publishes patch releases like
   1.20.6, 1.20.1, 1.20.4, and 1.21 as stubs — directories containing only
   version.json, because their game data is unchanged from the prior release. So
   get_data_path("1.20.6") returned a directory with no items.json /
   blocks.json / etc., and 5 parametrized tests soft-skipped to hide the gap.
2. common() was broken by the previous Pooch refactor. It still pointed at
   minecraft_data/data/data/ — a bundled-data directory that no longer exists in
   this repo — so it always raised FileNotFoundError. The test_common_data test
   caught that and skipped, masking the breakage.

Changes

• SUPPORTED_VERSIONS: ["1.20.6", "1.21.6"] → ["1.20.5", "1.21.6"].
  1.20.5 is the nearest preceding full-data version in the 1.20 series.
• __init__.py: common(): now reuses the Pooch-extracted data root via
  get_data_path(SUPPORTED_VERSIONS[0]). Common data is version-independent, so any
  supported version's data root works.
• test_common_data: removed the try/except FileNotFoundError +
  pytest.skip — it now asserts on the result directly.
• test_get_data_path_error_message: asserts "1.20.5" instead of "1.20.6"
  appears in the error.
• README: updated the supported-versions list and added a new section
  "Upstream data availability: full vs. stub versions" documenting which 1.20.x and
  1.21.x versions ship full data vs. stubs.
• Version: bumped 3.20.2 → 3.20.3.

Test impact

33 passed, 6 skipped → 39 passed, 0 skipped. Five skips disappear because
1.20.5 actually has the data the parametrized tests expect; the sixth disappears
because common() now works.

Not in scope

__call__ still has a # Fallback to bundled data if version not supported by Pooch branch that points at the same non-existent bundled directory. It's
unreachable in practice (any version not in SUPPORTED_VERSIONS raises
ValueError before the fallback runs in the current flow), but it's dead code
from the same refactor. Worth a follow-up.