Improve output generation step

[gaetanomatonti]

We should improve the performance of the output generation step.

Currently, if multiple input files are provided, we decode all of them in each sub-step of the output process. If the same file is referenced for different token types, we decode the same tree multiple times.

Ideally, when decoding a tree, we should cache it so that we don't repeat the decoding process (if needed).