Add vLLM ROCm 0.23 kernel tracing support

[irvineoy]

Summary

• replace the fixed vLLM ROCm v0.19.1 tracing registry with v0.23.0
• regenerate the image-pinned vLLM/AITER registry (841 traceable entries)
• pin generated Docker trace benchmark configs to the resolved registry image without modifying the source YAML
• update the default vLLM ROCm image, tracing documentation, and regression coverage

Validation

• pytest tests/test_kernel_tracing.py tests/test_kernel_tracing_cases.py tests/test_graders.py -q: 173 passed, 1615 skipped
• offline test suite (-m 'not live', excluding the unrelated existing dense-model registry assertion): 852 passed, 1615 skipped, 48 deselected
• representative v0.23.0 trace dry-run succeeded
• v0.23.0 Docker overlay import smoke test succeeded on GPU 0
• generated Kimi K2.6 trace config pins vllm/vllm-openai-rocm:v0.23.0 while leaving the original nightly config unchanged

Notes

The unfiltered suite has three unrelated environment/baseline failures: two live Codex/Cursor tests cannot write their user-state directories in the execution sandbox, and the existing Qwen3.5-27B model entry conflicts with test_dense_models_have_one_expert.