Fix gumbel

[alexander-telepov]

There are two commits, fixing corresponding issues:

• Gumbels shoul be summed with log-probablities of actions instead of probabilities inside gumbel-softmax trick;
• Non-existing attachment sites should have zero probablity to be choosen by network, i.e minus infinity log-probabilities.

[alexander-telepov]

g_ratio was changed from 1e-3 to 1.0 after replacement of probabilities to log-probabilities as input to gumbel_softmax function because I find out that in "freed pe" experiment after random exploration phase actor stacked in state, in which it sample only one action. In such state magnitude of gumbels multiplied by g_ratio was much smaller than log-probabilities, and actor was actually determenistic. After I changed g_ratio to 1.0 (in which function takes form same as original gumbel-softmax formula) actor doesn't stack in such states anymore.