fix(ci): purge sccache before test — stale L4 PTX fails on H100

sccache PVC holds CUDA kernel objects compiled for sm_89 (L4).
After moving the runner pool to H100 (sm_90), these produce
CUDA_ERROR_INVALID_PTX at runtime. Nuke the cache so candle
recompiles kernels for the correct compute capability.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.gitlab-ci.yml

@@ -219,6 +219,8 @@ test:
         fi
         echo "Waiting for Redis... ($i/30)"; sleep 1
       done
+    # Purge sccache — stale PTX from L4 (sm_89) causes CUDA_ERROR_INVALID_PTX on H100 (sm_90)
+    - rm -rf "${SCCACHE_DIR:?}"/* || true
     - sccache --zero-stats || true
     - cargo test --workspace --lib -- --skip model_loader::tests
     - sccache --show-stats || true