web: expose --gpu-memory-utilization in text-generation Deployment Options

[cswaney]

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

Context

vLLM's `--gpu-memory-utilization` (default 0.9) is the fraction of GPU memory vLLM claims for weights + KV cache. Lowering it is useful when co-locating with other GPU workloads on the same allocation, or when leaving headroom for CUDA graph capture and activation memory.

Users can pass it today via `launch_kwargs`, but it deserves a first-class control alongside the incoming `--max-model-len` field.

Scope

• Add a numeric input (0.0–1.0, float) to Deployment Options in `TextGenerationContainerOptionsForm.jsx`. Empty = vLLM default 0.9.
• Wire through `containerOptions.gpu_memory_utilization` in `text-generation.jsx`.
• In `buildContainerConfig`, append `--gpu-memory-utilization ` to `launch_kwargs` when set.
• Reflect the value in `ServiceSummary`.
• Tests: `TextGenerationContainerOptionsForm.test.jsx`, `web/src/lib/requests.test.js`.

Notes

• Likely lands after or alongside the `--max-model-len` PR; the `buildContainerConfig` merging pattern will be the same.