Add another epsilon

ml-agents/mlagents/trainers/torch/distributions.py

 
     def _mask_branch(self, logits: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
         raw_probs = torch.nn.functional.softmax(logits, dim=-1) * mask
-        normalized_probs = raw_probs / torch.sum(raw_probs, dim=-1).unsqueeze(-1)
+        normalized_probs = raw_probs / (
+            torch.sum(raw_probs, dim=-1).unsqueeze(-1) + EPSILON
+        )
         normalized_logits = torch.log(normalized_probs + EPSILON)
         return normalized_logits