Sharp description of local minima in the loss landscape of high-dimensional two-layer ReLU neural networks