[Bug] Render not always working for gym engine

[jelledouwe]

🐛 Bug

When using the gym engine, the human render screen does not always appear.

To Reproduce

Create object (for example pendulum-v1) with gym engine and single process backend.
Then call env.render("human") and train with sb3.

import eagerx
from eagerx_dcsc_setups.pendulum.objects import Pendulum
from eagerx_dcsc_setups.pendulum.envs import PendulumEnv
from eagerx.engines.openai_gym.engine import GymEngine
from eagerx.wrappers import Flatten
from eagerx.backends.single_process import SingleProcess

import stable_baselines3 as sb3
import gym.wrappers as w

if __name__ == "__main__":
    rate = 20

    # Create pendulum object
    pendulum = Pendulum.make(
        "pendulum",
        actuators=["u"],
        sensors=["x", "image"],
        states=["model_state", "max_speed", "length"],
        actuator_rate=rate,
        sensor_rate=rate,
    )

    # Create graph
    graph = eagerx.Graph.create()
    graph.add(pendulum)
    graph.connect(action="voltage", target=pendulum.actuators.u)
    graph.connect(source=pendulum.sensors.x, observation="angle_data")
    graph.render(source=pendulum.sensors.image, rate=rate)

    gym_engine = GymEngine.make(rate=rate, process=eagerx.ENVIRONMENT)

    backend = SingleProcess.make()

    # Create envs
    train_env = PendulumEnv(
        name="TrainEnv",
        rate=rate,
        graph=graph,
        engine=gym_engine,
        backend=backend,
    )
    train_env = w.rescale_action.RescaleAction(Flatten(train_env), min_action=-1.0, max_action=1.0)
    train_env.render("human")
    model = sb3.SAC("MlpPolicy", train_env, verbose=1, learning_rate=7e-4)
    model.learn(total_timesteps=5000)

Expected behavior

Render screen opening.
Note that a workaround is to do an extra reset before training.