Frontend simplification (4→2 tabs) + lazy imports for HF Spaces

Replace dual 8B with single 30B-A3B FP8 vision model

Reduce thinking model max_new_tokens to fix slow inference

Replace 30B MoE with dual 8B models (Thinking + Instruct)

Implement lazy model loading to prevent CUDA OOM on 4xL4 GPUs

Fix multi-GPU compatibility issues (6 locations)

Fix critical model implementations and add sample scenarios

Initial commit: FDAM AI Pipeline v4.0.1