Why context matters in VQA and Reasoning: Semantic interventions for VLM input modalities