DiffBio 0.1.7

calibrax 0.1.8 adoption

The classification loss is calibrax's softmax_cross_entropy(logits, labels) and diffbio.utils.training.cross_entropy_loss is gone; benchmark accuracy and macro-F1 come from calibrax with num_classes set to one past the largest label seen; the Lennard-Jones and compute-cost benchmarks time their steps through calibrax.profiling.time_calls; masked_value_loss and the chromatin-guidance loss reduce through calibrax's masked mse/mae (masked_value_loss no longer takes epsilon). Five documentation import blocks that 0.1.6 shipped broken are repaired. Requires calibrax 0.1.8.