feat: Encode Stan writer output using protobuf in C++

[riddell-stan]

Previously Python was used. Doing everything directly in C++ is much
faster and we eliminate the dependency on the Python protobuf library.

This version requires that the user install protobuf system-wide. Once
this solution is verified as working, we can remove this limitation by
copying the shared library into place.

Before this is merged:

• Clean up C++ to eliminate the use of stringstream, most uses are unnecessary.
• Copy the protobuf library and include files into the package, use rpath (see @riddell-stan 's example repo)

This version is about 36% faster on longer fits (lots of draws).

Closes #282
Closes #102

[riddell-stan]

This is not ready to be merged.

[riddell-stan]

@ahartikainen How do I tell gcc on Windows about a library and include directories I have installed with vcpkg? Here, in .appveyor.yml, I've used vcpkg install protobuf protobuf:x64-windows but then gcc complains about not finding the header files for protobuf.

Or is this not going to work because the shared libraries will not be compatible?

edit: add correction

[ahartikainen]

I think lib should on the PATH

[riddell-stan]

This contains a lot of useful information: microsoft/vcpkg#8577 . I'm still not sure precisely what to do. Testing is hard without a Windows computer. It seems like we're 95% of the way there.

[riddell-stan]

@ahartikainen Have a look.

Everything important happens in the first commit, 4cd2fff. The file httpstan/callbacks_writer_parser.py is removed. The parsing work in that file is moved into C++ files, in httpstan/queue_logger.hpp and httpstan/queue_writer.hpp.

Encoding the protobuf messages in C++ means we need to link with the protobuf shared library, which we have to compile and place into the Python package (so subsequent model compilations can use it).

[ahartikainen]

Looks good.

I wonder if there is a way to hack docker so it could be used for fast development?