Move generate_default_trace (and any other tools) to use same distribution layer as training

[mtrofin]

More reuse, common debuggability, etc. E.g. currently generate_default_trace in-worker failures are hard to debug.

[boomanaiden154]

It looks like generate_default_trace.py was handled by #222? It looks like the only other two tools left that aren't using the abstraction layer and just using the raw multiprocessing calls are generate_vocab.py and extract_ir.py (although the bulk of the actual multiprocessing code there is in extract_ir_lib.py).