Add packages/bench: single-task CUA model runner

[jarugupj]

Summary

• Adds a private @onkernel/cua-bench workspace — the first slice of a benchmark that runs CUA models on Kernel browsers and reports accuracy / cost / speed.
• runTask(modelRef, task) provisions a fresh Kernel browser, builds a CuaAgentHarness on the given model, runs the prompt, and tears the browser down. It captures wall-clock, turn count, and token totals (summed across every model call via harness events).
• spike.ts is a runnable one-shot: one HN task on anthropic:claude-opus-4-6.

Scope / what's deliberately left for follow-up

• success returns null — the accuracy judge (adopting an existing benchmark's scorer, e.g. Online-Mind2Web's) is the next step.
• costUsd is populated only when the provider reports a cost; the token×price conversion comes with the model price table.
• One task, one model, hardcoded — the task-set loader and multi-model fan-out are next.

How to run locally

npm install
export KERNEL_API_KEY=...
export ANTHROPIC_API_KEY=...
npm run spike --workspace @onkernel/cua-bench

Prints a TaskResult JSON with timing + token totals.

Test plan

• tsc -b typechecks clean
• spike resolves all imports and fails fast without KERNEL_API_KEY
• real run against claude-opus-4-6 produces wall-clock + token numbers