chore: sync text generation launcher with vLLM 0.26

[cswaney]

Summary

Two narrow fixes surfaced by the vLLM v0.26 audit.

Migrate `--model` to a positional argument. v0.26 emits a deprecation warning on every service start:

With `vllm serve`, you should provide the model as a positional argument or in a config file instead of via the `--model` option. The `--model` option will be removed in a future version.

Migrating now avoids a hard break on the next vLLM bump. Applies to `text_generation_slurm.sh` and both branches of `text_generation_local.sh` (docker + apptainer).

Remove the Disable Custom Kernels toggle. It has been a phantom option for a while: the toggle sets container state and appears in the request body, but neither the active `text_generation_*.sh` templates nor the backend translate it into a vLLM flag — only the `templates/deprecated/` copies mention `--disable-custom-kernels`. The toggle told users they were changing behavior; they were not. Cleans up UI, initial state, propTypes, tests, and dev fixtures.

Test plan

• `uv run just lint`
• `uv run just test` (969 pass)
• `npm run lint`
• `npm test` (558 pass)
• Manual: launch a text-generation service on Della, confirm the vLLM deprecation warning is gone from the container log.

Notes

• Sub-task of Update container versions #451. Follows the v0.26 bump (chore(lib): bump vllm/vllm-openai to v0.26.0 #523).
• Follow-up issues to file after this lands: expose `--max-model-len`, expose `--gpu-memory-utilization`, expose `thinking_token_budget`.

Closes #515

[claude]

Review

Small, well-scoped PR — both changes do exactly what the description says, and I verified there's no leftover state:

• No remaining disable_custom_kernels references outside templates/deprecated/ (checked Python backend services/base.py/asgi.py, and the frontend requests.js) — the field was truly dead everywhere except the deprecated templates, matching the PR's claim.
• buildContainerConfig in web/src/lib/requests.js is a generic passthrough, so dropping the field from the two default-options objects (text-generation.jsx, SidebarContainer propTypes) fully removes it from the request body — no orphaned reference.
• Positional-arg migration is applied consistently across all three templates (text_generation_local.sh docker + apptainer branches, text_generation_slurm.sh), with the positional model path placed before the flags in each, which matches vLLM's documented vllm serve MODEL [options] form.

Minor/non-blocking observations:

• --revision {{ container_config.revision }} is still passed as a flag alongside the positional /data/snapshots/{{ revision }} path — that duplication predates this PR, but since you're already touching every line around it, it might be worth a follow-up to confirm it's still needed now that the model path itself is revision-scoped.
• The "Manual: launch on Della" checkbox in the test plan is unchecked — worth confirming before merge that the deprecation warning is actually gone in a real container log, since that's the entire motivation for the --model change and templates aren't covered by the Python/JS test suites.
• lib/tests/dev/test_text_generation.py isn't part of the automated suite (manual requests-based smoke script), so the 969 pass / 558 pass counts don't actually exercise these template changes — the manual Della check is the only real verification here, reinforcing the point above.

No security, performance, or test-coverage concerns beyond the manual-verification gap noted above. Good cleanup.