fix(mcp): reap MCP child processes on process exit

[YOMXXX]

Summary

Fixes #1892. MCP child processes were orphaned when MiMoCode exited/restarted (reparented to PID 1, ~2 procs/~81MB leaked per restart; 22 shell+MCP pairs = 1.7GB observed).

Root cause: packages/opencode/src/mcp/index.ts already kills MCP server children + descendants in an addFinalizer (on Effect teardown). But the app's exit path does NOT dispose the instance — the TUI's onExit doesn't call instance.dispose(), and crash/SIGTERM never runs Effect teardown. So MCP children survived the process.

Fix: a module-level mcpChildPids registry + process-level exit handlers:

• Populated with transport.pid on each local MCP connect (success path only).
• Pruned in closeClient (disconnect/reconnect) so the exit handler can never SIGTERM a recycled PID.
• process.on("exit") + SIGTERM run reapMcpChildren() — a synchronous recursive pgrep -P kill of each tracked PID and its descendants (the graceful finalizer stays for Effect teardown).
• SIGINT deliberately NOT hooked so TUI Ctrl+C still uses graceful teardown.

Test Plan

• Real-process test: reapMcpChildren terminates a registered process; a two-level tree test covers a non-leaf PID; finally cleanup prevents leaked sleep processes.
• bun test test/mcp/reap.test.ts — 3 pass (10× stable); bun typecheck clean; MCP suite 115 pass no regression.

Notes

• Best-effort POSIX fix (uses pgrep); no-op/degraded on Windows.
• The SIGTERM handler deliberately exit(0) after reaping (masks the signal's non-zero code for supervisors); documented in code.