[rllib] Cleanup RNN support and make it work with multi-GPU optimizer (#2394)

Cleanup: TFPolicyGraph now automatically adds loss input entries for state_in_*, so that graph sub-classes don't need to worry about it.

Multi-GPU support:

Allow setting up model tower replicas with existing state input tensors

Truncate the per-device minibatch slices so that they are always a multiple of max_seq_len.

python/ray/rllib/agents/ppo/ppo.py

@@ -50,7 +50,7 @@ DEFAULT_CONFIG = with_common_config({
     "simple_optimizer": False,
     # Override model config
     "model": {
-        # Use LSTM model (note: requires simple optimizer for now).
+        # Whether to use LSTM model
         "use_lstm": False,
         # Max seq length for LSTM training.
         "max_seq_len": 20,