Reduce amount of memory required by mining strategies

E.g. in semi-hard mining we use cubic matrix and it consumes lots of memory.

quaterion/quaterion/loss/triplet_loss.py

Line 157 in f2be2a4

dists_tile = dists.repeat([batch_size_b, 1])

Due to this usage of memory train_cifar100_with_xbm.py crashes even on 3090Ti with "GPU out of memory" (after it exceeds start_iteration and starts computing xbm loss)

This is a known issue marked as TODO. Currently it allows to use XBM with smaller batches. We postponed it to be handled later on --you can assign it to me if you aren't working on it already.

Completed in #186