Ignore $ref siblings in supported JSON Schema drafts

[Kaiserunix]

Fixes #97.

Generating values for a non-recursive schema could fail with an unresolvable JSON pointer when a $ref had a sibling items schema. Resolve the referenced schema without merging sibling keywords, matching the supported Draft 4, 6, and 7 semantics. This also avoids treating conflicting sibling constraints as an unsatisfiable schema.

Regression tests cover both definitions and $defs locations across all three supported drafts, plus conflicting sibling constraints. The existing Draft 7 corpus case now declares its draft explicitly so its validation uses Draft 7 instead of the dependency's default newer draft.

Validation on Windows, CPython 3.12, with the repository's pinned test/check dependencies:

• Nine new regression cases fail on the unchanged implementation and pass with the fix.
• The original issue schema now generates 20 valid examples without changing the reproduction script.
• pytest -n 4 --durations=10 -k 'not test_can_generate_for_real_large_schema': 704 passed, 24 skipped, 52 xfailed, 2 xpassed; 100% coverage. These skips and expected-failure markers are existing suite policy, not added by this change.
• Shed for the changed Python files, Ruff, Black, JSON formatting, and git diff --check passed.
• Mypy reports ten errors in the unchanged _from_schema.py; the same command and environment against an independent checkout of the base commit produce the identical ten errors.

The separately scheduled slow corpus job and other OS/Python combinations were not run locally. This patch was prepared with AI assistance and independently reviewed.