Reduce number of forward passes during training for speed-up

[thorstenwagner]

During training, I do three independent forward passes through the network for the anchors, the positive and the negative volumes respectively (see https://github.com/MPI-Dortmund/tomotwin-cryoet/blob/2f12e793a6800dd58fe09573b75b801e26830b91/tomotwin/modules/training/torchtrainer.py#L626).

It is probably more efficient to concatenate them into one batch and do it in one pass.