[Bugfix][Rollout] Restore generation with vLLM 0.30

[0z5a]

Incremental review: Files changed.

Base: stack/vime-dev-h100-20260928 at c8be49d2b08a17d66870f01ded72f2a8b59191f8.

Restore startup and rollout endpoints with vLLM 0.30. Move the three serving CLI imports from entrypoints.openai to entrypoints.launchers, align the existing import fixture and set enable_scale_out=True for /inference/v1/generate.

Validation	Before	Candidate
Serving CLI	Import failure	Passed
Generation after import-only fix	HTTP 404	128 real requests completed on 2×H100
CLI/config/engine checks on H100 and H20	—	110 passed
Full Qwen3-4B run, 2×H20, TP=2	Startup blocked	Rollout → Adam update → full checkpoint → publication → normal exit

H20 timing	Candidate	Speed ratio
Generation	163.184 s	N/A: baseline cannot start
Actor training	88.415 s	N/A
Reported step	252.854 s	N/A
Whole process	768.372 s	N/A

Qwen3-4B revision 1cfa9a7208912126459214e8b04321603b3df60c, 16 prompts × 8 completions, 4096-token cap. Reward was 16/128, truncation 112/128 and gradient norm 0.1077845. A second process restored model, optimizer, scheduler and dataset cursor and completed five further online updates with finite nonzero gradients, checkpoint, publication and normal exit. Its warm median step was 190.601 s with KV cache increased from 16 to 32 GiB per card; that is not a code speedup.

The existing import fixture remains required by the namespace change; the added endpoint test was removed. Resident-actor optimization is reviewed separately in #2.