ci: increase compile-training pod timeout to 20 min

When all 8 per-service compile jobs fire simultaneously (e.g. on
Cargo.toml change), the single POP2-32C-128G node is fully loaded.
Training compile pod can't schedule until service jobs finish.
KUBERNETES_POLL_TIMEOUT=1200 gives it time to wait.

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

.gitlab-ci.yml

@@ -599,6 +599,9 @@ compile-training:
     # CUDA stubs from ci-builder image — compiles CUDA kernels without GPU hardware.
     # L4 GPU = compute capability 8.9 → candle-kernels generates sm_89 PTX
     CUDA_COMPUTE_CAP: "89"
+    # When all 8 service jobs fire (Cargo.toml change), the node is fully loaded.
+    # Increase pod scheduling timeout so training waits for service jobs to finish.
+    KUBERNETES_POLL_TIMEOUT: "1200"
   rules:
     - if: $CI_COMMIT_BRANCH == "main" && $CI_PIPELINE_SOURCE == "push"
       changes: