Require framework_version for framework estimators, models, etc.

[laurenyu]

Is your feature request related to a problem? Please describe.
Back when we were planning the migration of our TensorFlow and MXNet images to support script mode, we realized that upgrading the default framework version could unintentionally break existing workflows because frameworks can introduce breaking changes with new versions. This is why we stopped updating the default value for framework_version, and instead started showing warnings for when the default version was used.

Describe the solution you'd like
The proposal here is to make framework_version a mandatory parameter. The code changes required are fairly straightforward:

• Don't accept a Framework/FrameworkModel without a framework version anymore.
• Remove the default version for each framework, as well as the warning that displays when one does not specify framework_version.
• Move the "latest version" constant to the test files.

A corollary to requiring framework_version is to require py_version for frameworks with support for multiple Python versions.

One open question: should these parameters be required if an image URI is supplied? If not, how should we expose options like repacking a model, etc.?

edit: framework_version should be required even if the image URI is supplied because of our version-specific logic.

[laurenyu]

another thing the team talked about that could happen either as part of this task or (maybe more likely) #1464 is to create some kind of config file that keeps track of the various framework and Python versions that are supported