Fixed a bug in the random noise generation

There was a problem with the random noise generation. The problem came from the fact that the data was scaled by a factor of 1/data_ball which is a very big number by default (1e30). Instead of using Numpy's poisson function, we use the Python equivalent of ProxMatlab's PoissonRan. However, there were numerical precision issues. In Python, one exponential evaluation would give +inf, while it would give zero in Matlab. So, we chose a smaller scaling factor (currently, it is set to 1e8). This appears to work well. The JWST_CDRl_noise demo now works, as of July 18, 2020, in the master branch. Since then, several other JWST noise demos have been added, and everything matches ProxMatlab (using the same scaling factor).