ci(marketplace): decouple releases from GPU runner availability

[staging-devin-ai-integration]

Follow-up from PR #638 review (#638 (comment)).

marketplace-release.yml invokes the reusable marketplace-build.yml without build_cuda, which defaults to true, so build-marketplace-cuda (runs-on: [self-hosted, linux, x64, gpu]) always runs on release. continue-on-error: true handles step failures gracefully, but if no runner with those labels is registered the job stays queued indefinitely and downstream jobs (needs: [..., build-marketplace-cuda]) stall the whole release.

Possible mitigations:

• add a timeout-minutes to the CUDA job so a missing runner fails fast instead of queueing forever
• split the CUDA variant build into its own non-blocking workflow
• operational: ensure the GPU runner stays registered, or dispatch with build_cuda=false