Latent Traits and Cross-Task Transfer: Deconstructing Dataset Interactions in LLM Fine-tuning