gptj: fix abysmally slow postprocessor performance; don't read a file for each new batch

[git-bruh]

This wasn't helped by the fact that the "streaming" streams repeat the whole of the previous stream along with the new tokens...