Millimeter-wave (mmWave) cloud radio access networks (CRANs) provide new opportunities for accurate cooperative localization, in which large bandwidths, large antenna arrays, and increased densities of base stations allow for their unparalleled delay and angular resolution. Combining localization into communications and designing simultaneous localization and mapping (SLAM) algorithms are challenging problems. This study considers the joint position and velocity estimation and environment mapping problem in a three-dimensional mmWave CRAN architecture. We first embed cooperative localization into communications and establish the joint estimation and mapping model with hybrid delay and angle measurements. Then, we propose a closed-form weighted least square (WLS) solution for the joint estimation and mapping problems. The proposed WLS estimator is proven asymptotically unbiased and confirmed by simulations as effective in achieving the Cramer-Rao lower bound (CRLB) and the desired decimeter-level accuracy. Furthermore, we propose a WLS-Net-based SLAM algorithm by embedding neural networks into the proposed WLS estimators to replace the linear approximations. The solution possesses both powerful learning ability of the neural network and robustness of the proposed geometric model, and the ensemble learning is applied to further improve positioning accuracy. A public ray-tracing dataset is used in the simulations to test the performance of the WLS-Net-based SLAM algorithm, which is proven fast and effective in attaining the centimeter-level accuracy.