Fixes issue when training Boltz-1x with steering

[darianyang]

As noted by @jbderoo in a comment on PR #249 (and related to #243), training with fk_steering and guidance_update set to True results in a KeyError. This is because compute_constraint_features is set to False by default and couldn't be adjusted from within the training cfg file.

This minimal PR adds compute_constraint_features as an arg to the DataConfig and wrapper datasets in src/boltz/data/module/training.py so users could train with steering if they want to.

I'm not sure how useful/practical that would be, but now you can run it at least.

I also added the compute_constraint_features arg to the training cfg files, set to the default value of False since normally you wouldn't train with steering.