[rllib] Fix stats collection and some docs bugs since the refactoring (#2361)

* fix

* fix pbt example

* fix

* fix

* single thread by default

* vec

* fix

* fix

python/ray/rllib/optimizers/sync_samples_optimizer.py

@@ -17,12 +17,11 @@ class SyncSamplesOptimizer(PolicyOptimizer):
     model weights are then broadcast to all remote evaluators.
     """
 
-    def _init(self, batch_size=32):
+    def _init(self):
         self.update_weights_timer = TimerStat()
         self.sample_timer = TimerStat()
         self.grad_timer = TimerStat()
         self.throughput = RunningStat()
-        self.batch_size = batch_size
 
     def step(self):
         with self.update_weights_timer: