[rllib] Support eager execution in TF2

[ericl]

Describe the problem

Currently RLlib disables V2 behavior entirely. We should allow eager execution (at least for a subset of algorithms). This will be easier once #4788 is done.

One possibility is to still use graph mode, but allow eager in the loss function and model with:
https://www.tensorflow.org/api_docs/python/tf/py_function

[ericl]

It seems that with ModelV2 we can allow most policy methods, including the model, action distribution, and loss, to be executed eagerly. Currently trying to prototype this in the ModelV2 PR.

[ericl]

The long-term solution here is probably to have a separate EagerTFPolicy implementation of TFPolicy that implements the eager strategy. Policies can switch between that and the current TFPolicy. With the new build_tf_policy API, policy definitions should be compatible with either eager or graph execution.

I tried some variations of tf.py_function to allow eager within the existing TFPolicy, but there are some serious limitations there. In particular, variables referenced within tf.py_function aren't picked up when computing gradients.