`prediction_loader` for coupled eval

[elynnwu]

We have this feature for fme.ace.inference but not coupled. This is useful to compare things like inference/mean/weighted_mean_gen, where the current workaround is to manually add in the corresponding target from a different panel, but this is cumbersome and requires manually doing this for all variables. If we can have prediction_loader for coupled eval, we can simply do a run that's target data only and use it as a standalone wandb run to compare with.