[saraswatmanu]

Prepares the v2.0.0 release. Fitted results differ from v1.0.0, so refit your data rather than comparing networks across versions. The changelog lists the new defaults and how to restore them.

Main changes

• Faster phase 2: phase 1 is kept fixed and its calculations are cached, 10.9× faster on the mouse gastrulation example.
• Regularization fixed: the L1 and L2 penalties were computed on detached tensors, so they never affected training.
• TF expression: a new two-sided rule with a detection floor.
• Notebooks and docs: the same three notebooks, with saved outputs from the final fit, one shared config, and fuller explanations. New best practices page on how far results change between fits.
• Fresh installs: default genome downloads, the motif path and dependency pins are fixed.

Checks

• Install, notebooks and tests from a clean clone on Python 3.10 and 3.12, on small synthetic data on CPU.
• A fresh install on the gastrulation data with one GPU and shortened training: genome downloads and all three notebooks ran.

Issues

Fixes #15. Fixes #17. Fixes #19. Closes #20.

Addresses #13: fixed priors now stay frozen after reloading a checkpoint, which is a likely contributing cause. The reporters are asked to retest, so this issue stays open.