feature sets for other models, improve memory usage, cleanup old stuff

[jpata]

• follow-up to #62, implement feature set configuration for ParticleTransformer and LorentzNet
• unify model forward interfaces
• simplify ParticleTransformer by removing SequenceTrimmer and uu
• clean up unused / old / untested feature standardization and use_lifetime
• move beam stuff from the dataset to LorentzNet implementation, as it's an internal detail of the model
• disable creation of unused and confusing "outputs" dir
• rename cand_features -> cand_ParT_features for clarity, this is the default feature set as in https://arxiv.org/abs/2407.06788
• DataLoader loads data iteratively via row_groups, the RAM usage is now much lower, fixes #37
• re-enable binary classifier training, set max_cands=32 for binary classification #57 via the command-line configuration

[jpata]

Training works and gives losses that are expected with the PR

[jpata]

before and after this PR, LorentzNet has equivalent performance

while ParticleTransformer is now better (probably due to removing the SequenceTrimmer):

for SimpleDNN, we now revert back to the cand_ParT_features (in #62, the default was changed to kinematics+ParT+lifetimes), which is better: