Extend Bernoulli distribution to p=0 and p=1

[jucor]

So far, torch.bernoulli(p) requires 0 < p < 1. While this is strictly speaking mathematically correct, allowing in practice p = 0 and p = 1 to sample from the limit constant case is very convenient in many algorithms (eg. over-parametrization of a state space, where neurons can degenerate with a weight arbitrarily close to 0).

This tolerance for p=0 and p=1 is implemented in Matlab's Stat Toolbox's binornd(), as well as in the very useful Randraw package.

The attached patch implements this tolerance and documents it.

[clementfarabet]

It makes sense to me. I actually wrote an 'if then' somewhere recently to generate distributions with p=0 :-).

Everybody ok with that?

[andresy]

Sure!

[jucor]

Great, thanks for the reactivity :-) [I was too lazy to use the 'if then' used by Clement]