Sampling problems when no_model_error=True

[hdelongueville]

Similarly to Added small amount to pollution event size #139, a small non-zero value is needed for epsilon when no_model_error=True to avoid sampling problems and prevent NaNs in gradient of log-probability.

Instead of adding a small amount, taking the maximum may be "cleaner".

[hdelongueville]

CORRECTION: the issue was not the sampling but the fact that Yerror=0. The error was zero because mf_variability was used instead of mf_repeatability, while using averaging_period = ['4H', '4H', '4H', '4H']. The solution is to use averaging_period = [None, None, None, None] and averaging_error = False.