fix(vllm-4b): default to vLLM 0.23 on ROCm to stop the idle HSA spin

[kh0pper]

The 4b classifier bundle defaulted to rocm/vllm-dev:nightly_main_20260413 (vLLM 0.19.1rc). On that build, the first request leaves a native thread in VLLM::EngineCore spinning at 100% of a CPU core for as long as the container runs: the ROCm HSA runtime's event thread, whose AMDKFD_IOC_WAIT_EVENTS wait keeps returning immediately. This is vllm-project/vllm#41964, fixed in vLLM 0.21.0.

Change: the compose default moves to AMD's rocm/vllm:rocm7.14.1_rdna_ubuntu24.04_py3.14_pytorch_2.11_vllm_0.23.0, which is validated on gfx1151, and a two-line comment records why. A host that sets VLLM_ROCM_IMAGE in the bundle .env still overrides it.

Diagnosis on a gfx1151 host:

• py-spy showed every Python thread idle.
• gdb put the hot thread in libhsa-runtime64.
• A fresh container stayed idle until the first request, then spun.

Before/after on the same host, same checks (idle GPU):

	vLLM 0.19.1rc nightly	vLLM 0.23.0
latency p50 / p90, 20 short requests	0.201 / 0.203 s	0.181 / 0.182 s
answers on 5 fixed prompts	reference	identical
qwen3_xml tool call	structured get_weather call	identical
hottest engine thread 60 s after requests	99.9%	0.0%

Also checked:

• Every flag the compose passes is accepted by the new image.
• qwen3_xml is registered.
• 40 more requests showed no latency outliers (max 0.183 s).
• node --test tests/bundle-contract.test.js: 28/28 pass.

The new image is about 75 GB on disk, against about 11 GB for the old one.