Fix bugs in greedy exploration

Summary:
Greedy exploration had two bugs:
- available actions mask was ignored when sampling an exploratory action
- sampled exploratory action was not moved to proper device.

This Diff fixes those problems.

Reviewed By: jb3618columbia

Differential Revision: D54887996

fbshipit-source-id: b93958936c0d47b54cef6fa0966954b1c5c71c98

pearl/policy_learners/exploration_modules/common/epsilon_greedy_exploration.py

@@ -46,5 +46,8 @@ def act(
         if not isinstance(action_space, DiscreteActionSpace):
             raise TypeError("action space must be discrete")
         if random.random() < self.epsilon:
-            return torch.randint(action_space.n, (1,))
-        return exploit_action
+            return action_space.sample(action_availability_mask).to(
+                exploit_action.device
+            )
+        else:
+            return exploit_action