Instructions get unpickled in from-cache runs

This is a continuation of #67 (closed). The goal of that bug was to avoid unpickling the instructions when loading the kernel from cache. KernelExecutorBase still calls kernel.get_written_variables() which analyzes all the instructions.