Use exec to replace shell with code-server, keeping it as a direct child
of the Coder agent. This ensures code-server inherits the agent's mount
namespace where all Docker volume submounts are visible.

Previously, backgrounding with & caused code-server to orphan to PID 1
on script exit, which has a different mount view (only sees the base
/home/coder volume, not the nested .ssh, .config, etc. submounts).