fix: Explicitly exit CLI to prevent hanging subprocesses

[veracioux]

Reproduced using MCP config from: #1717

The easiest way to reproduce (and verify the fix) is to run opencode run hi.

Investigation:

• The MCP servers from the issue are run via docker run
• When opencode exits, the ai library sends SIGTERM to all child MCP processes that use stdio transports
• The problem is that the child processes are the docker run client, and SIGTERM isn't properly delivered, unless docker run --init was used. So the subprocesses stick around, preventing opencode from exiting. I verified this with a minimalistic subprocess-spawning bun program and strace
• What does work though, is killing the opencode process itself using process.exit, in which case SIGTERM is delivered correctly
• Fix: I added a process.exit() when the CLI is done with everything.

NOTE: Issue also happens in (some, maybe all) MCP servers wrapped by shell scripts, which this also fixes.

Closes #1717.
Maybe also #1810.
Maybe also #2406.

[veracioux]

Rebased to remove unintended extra commit.