model-serving: a Muse Glimmer 30B preset once the runtime carries vLLM 0.28 or later #597

Problem

The signed model image for Meta's Muse Glimmer 30B (RedHatAI FP8-block repack, Apache 2.0) exists, but the platform's llm-d runtime image (llm-d-cuda v0.8.0) carries vLLM 0.23.0, and the model's architecture and its muse_glimmer tool and reasoning parsers first appear in vLLM v0.28.0 (vllm-project/vllm#51655). No llm-d release carries 0.28.0 yet (v0.9.0 = vLLM 0.26.0). A preset today would never start.

Proposed solution

• When an llm-d release with vLLM ≥ 0.28.0 exists and its image is mirrored and signed on gsoci: bump the runtime in kserve-runtime-configs, then add the preset muse-glimmer-30b on oci://gsoci.azurecr.io/giantswarm/models/muse-glimmer-30b-fp8:1deb4641ff84 (32 GiB FP8, one 48 GB GPU, contextLength: 32768, muse_glimmer parsers, image 1).
• Alternatively, a per-preset runtime image override on a signed mirror of a vLLM ≥ 0.28 image, if that mechanism ships first.

Acceptance criteria

• The runtime serves MuseGlimmerForConditionalGeneration (a serve in the lab or on a test installation reaches Ready and answers a tool call).
• The preset ships and verify-preset-weights passes.

Follows #591.