Feature Request: Add partial/window screenshot support to cut latency & token usage

First of all, huge thanks for releasing the computer-use-demo. We genuinely enjoy using it and are excited to roll it out more intensively in our
internal testing environment.

**The Problem:**
Inside a Ubuntu 22.04 virtual machine (but the same happens when the demo is run via Docker Compose), I’m using the computer-use-demo agent to analyse a single app window.
The built-in screenshot() helper always captures the entire desktop, even when the target window covers only ~25 % of the screen. Converting that full-screen PNG to base-64 and streaming it to Claude:
• adds ~3–4 s to every analysis cycle
• inflates prompt size by 3 × – 4 × (unnecessary token cost)

**Desired behaviour**
• Partial (region) capture – supply x, y, width, height or a bounding-box tuple.
• Active-window capture – supply the window name.
Either mode would trim the image to the relevant context, speeding up cycles and shrinking the base-64 payload dramatically.

**Alternative Solution**
If adding this directly isn’t on the roadmap, could you please:
• Document the recommended extension points for the computer tool, especially around customizing the screenshot() helper, or
• Expose a simple plug-in hook (e.g., ComputerMixin or env-var override) so we can swap in a custom screenshot command without patching the library.

Thanks again for the awesome tool!