Text this: Benchmarking Visual Feature Representations for LiDAR-Inertial-Visual Odometry Under Challenging Conditions