Extension + examples: migrate to document.modelContext, forward cancellation end to end

[kashishhora]

Summary

Stacked on #54. Moves the bridge extension and the examples off navigator.modelContextTesting (removed in Chrome 152.0.7940.0) and forwards MCP client cancellation all the way to the tool handler. Base is the 1.1.0 branch; GitHub will retarget to main once that merges.

• extension/src/content-main.ts — a small adapter over document.modelContext.getTools() / executeTool() (native Chrome or the 1.1.0 polyfill), falling back to modelContextTesting for pages still on webmcp-react <=1.0.0. It re-detects the page object by identity: webmcp-react installs a fresh modelContext whenever its provider remounts, so a held reference would go stale. inputSchema is normalized to a JSON string on the wire whether Chrome hands back a string (<=153) or an object (154+).
• Cancellation, end to end — MCP notifications/cancelled → SDK extra.signal → WebSocket CANCEL_TOOL → background → content scripts → executeTool(tool, args, { signal }). The 30s call timeout now sends CANCEL_TOOL too. Late results for a cancelled call are dropped at both the background and the server.
• packages/webmcp-server/bin regenerated so the published server carries the cancel path.
• examples/native-harness — the self-test drives getTools() / executeTool() and adds probes for string-vs-object inputSchema, the execution signal, a mid-flight abort, and unregister-during-execution.
• examples/playground — DevPanel on the consumer API with a Cancel button, plus a slow_hint tool to try it on.
• Extension README/PRIVACY and a CHANGELOG entry under Unreleased.

Test plan

• pnpm typecheck:extension / pnpm build:extension — the committed bin rebuilds byte-identical
• Playground and native-harness builds; library suite still 205 passing
• DevPanel: cancel slow_hint mid-flight against the polyfill — the result row shows the abort, other tools keep working
• Native Chrome Canary: harness self-test plus one extension list → execute → cancel round-trip. Not run yet — since native support is the point of this PR, I'd hold merge until it is.