[Docs] Prepare Engine best practices doc

[MichaelBroughton]

With all the information we have floating around with how to best use Engine + TFQ we might want to add a page like this: https://cirq.readthedocs.io/en/latest/google/best_practices.html

but for TFQ soon. Major points we need to make are:

1. Minimizing calls to the service where possible.
2. Controlling parallelism at op level with tf.CriticalSection or tf.control_dependencies
3. Controlling parallelism in layers and differentiators with (low_latency boolean) [Yet to be implemented].
4. Use of SampledExpectation where possible with understanding of how samples_per_term will scale the cost.
5. Encourage tf model check-pointing in case experiments need to be interrupted and resumed across time slots.
6. All best practices from here https://cirq.readthedocs.io/en/latest/google/best_practices.html as well.

I'm sure I've missed a few things so we should definitely work to add to this list over time.