Very valid question.
I too had the same question and tried the code with both Normal and Uniform noise.
With uniform noise MNIST images are are not generated clearly even after 100 epochs .
With normal distribution I did get good MNIST images generated from the Generator.
The reason why normally distributed noise works better than uniformly distributed noise may be due to the reason that normal distribution is the most common distribution.
Approximation also improves when we increase the size of the simple random samples that are used to produce normal sampling distribution
Please let me know if the explanation helps