Negative Samples Mining Matters: Reconsidering Hyperspectral Image Classification With Contrastive Learning