Replace Docker support with support for static data being provided to primitives

Supporting primitives using Docker seems to introduce complexity everywhere: it is hard for TA1 authors to write such primitives and it is hard to run them. There might be issues efficiently sending data from TA2 to primitives. And it is unclear how to pickle such primitives.

Current main use case known for using Docker container seems to be that a primitive needs a lot of data (e.g., a pretrained model) which would maybe make the base image really large. So a primitive could put this into a Docker image and so only those who choose to run the primitive would have to fetch that image and get this data.

But this could be addressed easier: by having primitive declare in its metadata that it needs some static data available during runtime. Then runtime can expose this data and provide paths to data to a primitive. Managing static files is much easier than whole Docker containers.