ViT Multi-Head Self-Attention

Implement Vision Transformer multi-head self-attention with combined QKV projection, scaled softmax, and output projection.