web: expose --max-model-len in text-generation Deployment Options

[cswaney]

Surfaced during the vLLM v0.26 audit (#515).

Context

vLLM's `--max-model-len` overrides the model's advertised context length. It is the most common OOM knob on Della — dropping context length from 131072 (Gemma 3's default) to, say, 32768 typically halves KV-cache memory and lets larger models fit or leaves headroom for concurrent requests.

Users can pass it today via the free-form `launch_kwargs` field (or CLI `launch_kwargs`), but it deserves a first-class control given how often it's needed.

Scope

• Add a numeric input to the Deployment Options section in `TextGenerationContainerOptionsForm.jsx`. Empty = use model default (unchanged behavior).
• Wire through `containerOptions.max_model_len` in `text-generation.jsx` initial state.
• In `buildContainerConfig` (which currently only translates `disable_thinking`), append `--max-model-len ` to `launch_kwargs` when set. Order matters if `disable_thinking` is also set — concatenate rather than overwrite.
• Reflect the value in `ServiceSummary` so users can see what a running service is configured with.
• Tests: `TextGenerationContainerOptionsForm.test.jsx`, `web/src/lib/requests.test.js` for the `buildContainerConfig` merging behavior.

Non-goals

• Enforcing a minimum/maximum. vLLM validates the value at startup and reports a clear error if the request exceeds the model's true context length.
• Auto-suggesting based on GPU memory. Nice-to-have, but the calculation involves dtype and TP shape; leave it to the user for now.