[Runtime] Add Qwen3-32B, Qwen3-Embedding-8B, and Qwen3Guard-Gen-8B vLLM runtimes

[Juno13340]

What this PR does

Adds OME configuration for serving three Qwen3 models on vLLM:

• Adds the qwen3guard-gen-8b ClusterBaseModel pointing to hf://Qwen/Qwen3Guard-Gen-8B, and completes the qwen3-embedding-8b ClusterBaseModel metadata (architecture / format / framework / parameter size).

• Adds the vllm-qwen3-32b ClusterServingRuntime with SMG router + vLLM settings for Qwen3ForCausalLM, 4-way tensor parallelism, 32K context, chunked prefill, Qwen3 reasoning parsing, and Hermes tool-call parsing.

• Adds the vllm-qwen3-embedding-8b ClusterServingRuntime (embedding/pooling via --runner pooling, Qwen3ForCausalLM, TP=1).

• Adds the vllm-qwen3guard-gen-8b ClusterServingRuntime (text-generation guard, Qwen3ForCausalLM, TP=1).

• Registers the models and runtimes in the kustomizations.

• Adds sample InferenceServices for the Qwen namespaces.

All runtimes use docker.io/vllm/vllm-openai:v0.20.0. Framework versions are pinned to each model's upstream config.json transformers_version (32B: 4.51.0, Embedding-8B: 4.51.2, Guard-Gen-8B: 4.51.1) to match the runtime selector's Equal comparison.

Why we need it

Enables serving for Qwen3-32B (chat), Qwen3-Embedding-8B (embeddings), and Qwen3Guard-Gen-8B (content moderation).

Fixes #

How to test

Validated each engine locally on an 8×A100 host via standalone docker run against vllm/vllm-openai:v0.20.0:

• Qwen3-32B — TP=4, /v1/chat/completions returns expected output.
• Qwen3-Embedding-8B — --runner pooling, /v1/embeddings returns embedding vectors.
• Qwen3Guard-Gen-8B — TP=1, /v1/chat/completions returns expected output.

kubectl kustomize config/models and kubectl kustomize config/runtimes both build cleanly.

Checklist

• Tests added/updated (if applicable)
• Docs updated (if applicable)
• make test passes locally