[sreeram-11]

Fixes the nightly OpenVINO build, which has been failing since upstream moved to OpenVINO 2026.4.1 on Oct 3, + two release-safety gaps found along the way.

1. OpenVINO build fixed:
   • The job now reads the OpenVINO version from upstream's workflow (at the pinned commit) instead of our hardcoded 2026.0, which lacks ov::ProfilingInfo::start_time.
   • It builds on both Ubuntu 22.04 and 24.04 as separate release families.
   • Artifact names change to llama-<tag>-bin-ubuntu-{22,24}-openvino-<version>-x64.tar.gz, and the release-notes links are now generated from the files actually published.
   • libze_loader.so.1 allowed in the ldd check: OpenVINO 2026.4 ships an NPU library that links it. Like libOpenCL.so.1, it comes from the user's Intel driver.
2. sm_121 is now verified: it was built but missing from the release completeness check, so a failed sm_121 build would have published an incomplete CUDA set.
3. New check-cmake-flags job: fails if any -DGGML_*/-DLLAMA_* flag we pass no longer exists (removed or renamed) in upstream's CMake files, and warns when our shared CMAKE_ARGS differ from upstream's. Nothing depends on it, so a failure turns the run red without blocking builds or the release.
   • Dropped -DGGML_HIP_ROCWMMA_FATTN=ON: upstream removed this option on Jul 24 (HIP: remove rocWMMA FlashAttention ggml-org/llama.cpp#26046), so it has been a silent no-op in both ROCm builds since then.
4. Package validation reports every problem at once: the RPATH/ldd checks in the OpenVINO and Linux CUDA jobs now go through every binary and fail at the end with a count, instead of stopping at the first bad file.

Notes:

• The drift check will warn about -DLLAMA_CONNECT=ON. I left it out because it downloads an unpinned latest binary at build time.
• PR runs don't run the release job, so the sm_121 check, the new OpenVINO families and the release notes only get exercised on the first nightly. I tested those steps locally against fake release directories.
• Supersedes Fix OpenVINO build and track upstream drift #28.

[sreeram-11]

The first attempt of the CI workflow run for commit 178ee6a (run #1) was failing on the CUDA and ROCm build jobs.

The cause was that this run pinned upstream commit 280141e6d4 (b11841), which came from upstream PR ggml-org/llama.cpp#30804. That PR broke upstream master for every CUDA and ROCm build, including upstream's own CI (cuda, hip, rocm, gpu-cuda and server-cuda jobs are all red).

I fixed upstream by ggml-org/llama.cpp#31053, which is now pulled in here to resolve the issue.