Make the RandomHyperTreeGridSource mask distribution more even with large masks

This brings a distribution that favorizes the unmasking of the small cells. The old one led to having only very large cells unmasked with a MaskedFraction > 0.5.

To do this we added multiple shuffles to remove the biais on our algorithm, which tends to mask the first trees it processes. We also used the level of the node to ponderate weither or not it should be masked, now the deeper the node is (the smaller it is), the less the chances are of it to being masked.

Also we had to change the seed depending on the piece to avoid having observable structures repeating in distributed mode.

Also now teh masking is consistent from increasing / decreasing the MaskedFraction value.

It still has a drawback (which was also present in the previous implementation), which is that due to the error margin used : the value 1 for the MaskedFraction won't mask the whole HTG, it is still possible to mask the whole HTG but it requires a small error margin (which depends on the parameters).