[RFC] Python Dependencies in TVM CI Containers

One potential idea is we can first start with l1, which is simpler, then see if there is need to move to l4. Personally I don’t think we really need a stronger consistency of packages in CI spectrum, sometimes we might want to diversify it a bit(e.g. use different compiler such as clang/gcc). We are also producing different tlcpack binary packages for different system, so we could derive the cpu package from ci-cpu and gpu package from ci-gpu.

Additionally, most of the dependencies that are relevant should have already captured by the constraint.txt e.g. tensorflow==1.6 so there is not a lot of room to pinning down further.