VoxGenesis: Unsupervised Discovery of Latent Speaker Manifold for Speech Synthesis