🔨 replace Netty server engine with CIO on desktop

[guiyanakuang]

Summary

Part of #4908 (thread count growing over long uptime).

The desktop sync server and the local MCP server were the only Netty users in the app. Netty keeps three event loop groups sized for a server host (connection = parallelism/2+1, worker = parallelism/2+1, call = parallelism), starts each event loop thread lazily on first use, and never releases it until shutdown. On an 18-core machine that is up to 38 permanently parked eventLoopGroupProxy-* threads that only appear to "grow" over days because they start on demand.

This PR swaps both servers to Ktor's CIO engine and removes the Netty dependency entirely:

• DesktopServerFactory / DesktopPasteServer / DesktopNetworkModule now use CIOApplicationEngine.
• DesktopMcpServer uses embeddedServer(CIO, ...); the MCP SDK's mcp {} plugin is engine agnostic.
• ktor-server-netty is dropped from the version catalog and the desktop dependencies, along with the -Dio.netty.maxDirectMemory JVM flag. The whole io.netty:* 4.2.x stack disappears from the runtime classpath.
• The Netty bootstrap options are gone: Ktor's network layer already sets TCP_NODELAY on accepted sockets, and the 30 s WebSocket ping already covers keep-alive.

CIO has no dedicated threads. The selector loop and request handling run on Dispatchers.IO, whose idle workers are reclaimed after 60 s, so the server no longer contributes a fixed thread pool at all. The CLI server has been running on CIO since it was introduced, and the commonMain routing is engine agnostic (the mobile apps cannot use Netty), so no routing or protocol code changes.

Verified from Ktor 3.5.2 sources that CIO's connectionIdleTimeoutSeconds (45 s) only applies while waiting between HTTP responses on a keep-alive connection; after a WebSocket upgrade the pipeline closes its writer channel and hands the connection over, so long-lived sync sockets are unaffected.

Test plan

• ./gradlew app:desktopTest (fast tier): 1680 tests, 0 failures
• ./gradlew app:desktopTest -PintegrationTests --tests SyncIntegrationTest (3 servers, full pairing + sync in one JVM): 15/15
• DesktopPasteServerTest port-in-use fallback path still passes
• app:dependencies --configuration desktopRuntimeClasspath contains no Netty artifacts
• Real-machine: two devices, large file transfer throughput and a WebSocket sync session idle for more than 45 s
• Real-machine: thread count (ps -M <pid> | wc -l) after several hours compared with 2.2.0