fix: use CLIPImageProcessor and remove direct 384x384 size override to fix prediction discrepancy

feat: performance overhaul, client-side report, result cache, streaming cards, UX polish

fix: correct layer indexing, update report generation method, and implement Captum hook cleanup to prevent model contamination

wtf

Use get_vit_encoder_layer for robust layer lookup across transformers versions

Robust layer extraction for GradCAM

Fix GradCAM target layer reference for transformers ViT model

Initial commit for XAI Space