chore(lib): bump vllm/vllm-openai to v0.26 or v0.27

[cswaney]

Sub-task of #451.

Bump the pinned vLLM image in `images.py` from `v0.20.0` to `v0.26.0`. `v0.27`+ is unnecessary unless we start using Qwen 3.5.

Draft PR is open at #523. Check items off here as smoke tests pass; mark #523 "Ready for review" once the matrix is green.

Expected Blackfish-side changes

Blackfish's launch template only uses mainline vLLM engine flags — `--model`, `--port`, `--revision`, `--trust-remote-code`, `--tensor-parallel-size`. None appear in vLLM's deprecation/removal lists between v0.21 and v0.26. The only other Blackfish-owned wiring is the Disable Thinking toggle (`--default-chat-template-kwargs '{"enable_thinking": false, "thinking": false}'`), whose flag name and JSON key shape were not touched in the release notes we scanned. Expected diff: pin only, plus incidental test/fixture bumps.

Landmines to watch for during the smoke test:

• v0.21 deprecates transformers v4 (image dependency change).
• v0.24 stops setting `CUDA_VISIBLE_DEVICES` internally; we drive device count via `--tensor-parallel-size` so this should be transparent.
• v0.25 removes legacy PagedAttention; V1/MRv2 backends are the standard path.
• HPC-specific: watch for CUDA driver mismatches on Della compute nodes and `apptainer pull` issues on the multi-arch manifest.

Runtime override for testing

No code change needed to pilot the bump:

```shell
BLACKFISH_TEXT_GENERATION_IMAGE=vllm/vllm-openai:v0.26.0 blackfish start
```

Stage the SIF on Della once:

```shell
apptainer pull vllm-openai_v0.26.0.sif docker://vllm/vllm-openai:v0.26.0
mv vllm-openai_v0.26.0.sif <cache_dir>/images/
```

Model matrix

For each: launch the service, confirm HEALTHY, hit one completion request through the tunnel.

Regression targets — popular models we already run

• google/gemma-3-4b-it — Gemma 3 confirmed working
• meta-llama/Llama-3.3-70B-Instruct — flagship dense; needs TP≥4

New-arch wins — models that were hit-and-miss or failed on v0.20.0

• google/gemma-4-12b-it — Gemma 4 support landed v0.21+; already confirmed working
• deepseek-ai/DeepSeek-OCR-2 — already confirmed working
• Qwen/Qwen3.6-27B — Qwen 3.6 confirmed working
• Qwen/Qwen3.6-35B-A3B
• openai/gpt-oss-20b — arch landed v0.24 (model loads—known vocab issue persists)
• meta-models/Muse-Glimmer-30B (failed)

Parameter checklist

Launch-time

The command is `apptainer run --model ... --port ... --revision ... --trust-remote-code --tensor-parallel-size {launch_kwargs}`. Testing the fixed flags happens implicitly with every launch above; the interesting bits are:

• Multi-GPU — --tensor-parallel-size still works (implicitly exercised by Llama-3.3-70B and Qwen/Qwen3.6-27B)
• Disable Thinking — launch Qwen/Qwen2.5-72B-Instruct with the Disable Thinking checkbox on. Container log has no "unknown option" warning for --default-chat-template-kwargs; a subsequent completion request comes back without <think>…</think> in the response.

Request-time (OpenAI-compat API served by vLLM)

Against any healthy service (TinyLlama is the cheapest), one chat completion request exercising several sampling params at once:

• Body: {temperature: 0.7, max_tokens: 64, stop: [\"\\n\\n\"], stream: true, seed: 42, frequency_penalty: 0.3, presence_penalty: 0.3}. Should return normally (no 400). Repeat with seed: 42 and verify deterministic output.

Land the bump

The draft PR (#523) already handles the pin update and the load-bearing test change. Once the checklist above is green, mark #523 "Ready for review" and merge.

Post-merge

• chore(web): sync text generation launcher params with the bumped vLLM version #515: verify the launcher param audit — no vLLM args added/renamed/removed that need frontend exposure.
• Note the bump in release notes / docs if a release cut is imminent.