[rllib] Fix use_lstm option when using custom model with dict space (#3368)

## What do these changes do?

This passes in the right obs space to the lstm model wrapper, so that it doesn't attempt to un-flatten the already processed dict observation.

## Related issue number

Closes https://github.com/ray-project/ray/issues/3367

test/jenkins_tests/multi_node_tests/test_rllib_eval.sh

@@ -1,23 +0,0 @@
-#!/bin/sh
-
-# TODO: Test AC3
-ALGS='DQN PPO'
-GYM_ENV='CartPole-v0'
-
-for ALG in $ALGS
-do
-  EXPERIMENT_NAME=$GYM_ENV'_'$ALG
-  python /ray/python/ray/rllib/train.py --run $ALG --env $GYM_ENV \
-    --stop '{"training_iteration": 2}' --experiment-name $EXPERIMENT_NAME \
-    --checkpoint-freq 1
-
-  EXPERIMENT_PATH='/tmp/ray/'$EXPERIMENT_NAME
-  CHECKPOINT_FOLDER=$(ls $EXPERIMENT_PATH)
-  CHECKPOINT=$EXPERIMENT_PATH'/'$CHECKPOINT_FOLDER'/checkpoint-1'
-
-  python /ray/python/ray/rllib/eval.py $CHECKPOINT --run $ALG \
-    --env $GYM_ENV --no-render
-
-  # Clean up
-  rm -rf $EXPERIMENT_PATH
-done

test/jenkins_tests/run_multi_node_tests.sh

@@ -234,9 +234,6 @@ docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
     --stop '{"training_iteration": 2}' \
     --config '{"num_workers": 2, "optimizer": {"num_replay_buffer_shards": 1}, "learning_starts": 100, "min_iter_time_s": 1}'
 
-docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
-    sh /ray/test/jenkins_tests/multi_node_tests/test_rllib_eval.sh
-
 docker run --rm --shm-size=10G --memory=10G $DOCKER_SHA \
     python /ray/python/ray/rllib/test/test_local.py