Record container images in test-metadata.json and fix embedding vllm-bench path for RHAIIS images

[maryamtahhan]

Problem:

• Embedding test-metadata.json only records vllm_version from the server health check, not the container image tag. This made it impossible to tell from results alone whether RHAIIS 3.5.0 GA vs RC was used.
• LLM benchmarks already record vllm_container_image and guidellm_container_image; embedding, audio, and MTEB do not.
• When VLLM_BENCH_CONTAINER_IMAGE is set to a RHAIIS image, embedding benchmarks fail with crun: executable file /opt/venv/bin/vllm not found because baseline.yml / latency.yml hardcode the upstream image path.

Proposed changes:

1. Add vllm_container_image (from effective_container_image) to embedding, audio, and MTEB metadata
2. Add loadgen_container_image where a separate loadgen container is used
3. Standardize field names across all playbooks (match llm-benchmark-auto.yml)
4. Fix embedding bench command to use --entrypoint vllm instead of /opt/venv/bin/vllm
5. Update dashboard/conversion tooling to display the new fields

Context: Successful RHAIIS 3.5.0 embedding runs used split images: RHAIIS server + upstream vllm-openai-cpu:v0.25.1 bench client. The reported error occurs when RHAIIS is used for both.