ci: download GGUF models once upfront, restore-only everywhere (DRY the pipeline)

Replaces the per-job model download logic (duplicated across all 6 Java test
jobs) with a single upfront download-models job; every test job now only
restores the shared cache.

- New download-models job (needs: startgate) is the single source of the
  download logic: the ten curl steps + validate-models.sh live only there. It
  restores the shared cache (key gguf-models-v1) or, on a cold cache, downloads
  the full ~5 GB set, validates, and saves it.
- All 6 Java test jobs (Linux, 3x macOS, 2x Windows) drop their duplicated
  download blocks and now needs: download-models + restore-only the cache,
  keeping validate-models.{sh,bat} as a per-job integrity guard. Removes the
  cold-start save race; net -91 lines in publish.yml.
- The llama-langchain4j integration job now needs [crosscompile-linux-x86_64,
  download-models] instead of chaining behind test-java-linux-x86_64, so it runs
  in parallel off the guaranteed-populated cache.
- CLAUDE.md updated: the model-cache section documents the upfront download-models
  job (replacing the old "save race" description) and the integration job wiring.

YAML validated; structural counts verified (10 curl steps only in download-models;
7 jobs restore the cache; 7 validate steps).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE

Author: claude