Device selection `train_eval_multithreaded`

Hi,

I am a bit confused about the decide selection if running with `train_eval_multithreaded`. In my set up I have two GPUs and in training I use `pmap` and for evaluation I only use `jit`. How are the resources shared in this case? Since I do not explicitly select any devices, I think that here device 0 will do training and evaluation simultaneously. Is this how `jaxline` is intended/designed or am I missing something?

Thanks!

Start at the train_eval_multithreaded entry point and trace how pmap-based training and jit-based evaluation select devices when two GPUs are available. Confirm whether device 0 can run both phases concurrently, then document the intended resource-sharing behavior and any explicit selection needed.

Written by the indexing model from the issue text.