Is your feature request related to a problem? Please describe.
Currently we use np.random.rand to perform draws from prior which may exhibit clustering, thereby potentially biasing SBI training if the Prior.draw function is used for training data generation.

Describe the solution you'd like
Using Latin Hypercube Sampling ensures that the samples are spread more evenly across the entire domain, making it a better choice for high-dimensional spaces when compared to simple random sampling.