Glance-or-Gaze: Incentivizing LMMs to Adaptively Focus Search via Reinforcement Learning