[BUG]: Remove Reverb sampler from trainer tf.function.

[DriesSmit]

Describe the bug The dataset sampler should still be removed from the MADDPG/MAD4PG tf.function training steps and placed outside. If the sampler is inside tf.function it can possibly sample random noise data, because tf.function messes with its stop calculations.

[jcformanek]

In my experience, having the reverb sample inside tf.function is only a problem when you use a queue. So I expect it to be fine when using a regular replay buffer like in MADQN and MADDPG. Am I wrong? @DriesSmit

[DriesSmit]

Closing all TF issues as we are depreciating our TF systems.