FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model