Investigate indexing and micro-optimizations in vision encoder forward pass

Investigate potential micro-optimizations in the vision encoder forward pass.

This includes evaluating whether pre-allocating lists and using indexing (instead of append) brings any measurable performance benefits in methods that are executed frequently.

Additionally, review naming related to multimodal feature processing. The current name `feature_fusion` may be misleading since the implementation performs simple concatenation followed by a linear projection. Consider renaming it to `feature_adapter` or a similar term.