Fix doc build by removing tensorflow dependency (#2610)

Currently the doc builds are failing due to running out of memory when pip installing tensorflow.
I investigated fixing this problem by upgrading to the paid version of readthedocs, and found out that:

1. They do not have a way to just turn your free project into a paid project. You need to configure a new project.
2. Paid projects have *less* (**LESS!**) control over the domain name than the free projects (the URL would be prefixed with "quantum-lib-", also the top-level domain would differ; I strongly suspect we would not be able to use a custom domain that was not owned by us such as the existing docs location).
3. The support staff does not think that this is a situation that needs to be fixed.

So as a fix I'm killing off the TPU translation code from contrib and removing the tensorflow dependency.
I will also look into using other documentation services.

The tpu code is in contrib, so it does not need to be deprecated. The tpu code was still slower than non-tpu simulation due to the large compile times.

As a side benefit, this will likely speed up our builds.

cirq/contrib/contrib-requirements.txt

@@ -2,4 +2,3 @@
 
 ply>=3.4
 pylatex~=1.3.0
-tensorflow~=1.14