v0.3.0

This is a major update with breaking changes that introduce both new models and a new API.

• We introduce new "relaxed" updates that drop the mean-field assumption in the posterior and prediction update, resulting in uncertainties flowing both top-down and bottom-up via value coupling. This is of particular interest for deep networks.
• The DeepNetwork class now uses Equinox internally and scans over the ayer axis, allowing to train deeper model without hitting the compilation burden.
• All main classes let the user control volatility updates, mean-field updates, and clipping threshold. update_type is now volatility_updates.

What's Changed

• add relaxed precision updates by @LegrandNico in #384
• add relaxed posterior updates by @LegrandNico in #385
• use docformatter to format docstrings by @LegrandNico in #386
• refactor relaxed posterior updates for all three backends by @LegrandNico in #387
• Use Equinox internally for DeepNetwork by @LegrandNico in #388
• Set unbounded updates as default across backends by @LegrandNico in #389
• depreciate the distribution module and the HGF class by @LegrandNico in #390
• control volatility and mean-field updates for Network by @LegrandNico in #391
• new eHGF updates aligned with Matlab by @LegrandNico in #392
• set precision clipping values as parameters by @LegrandNico in #393
• all functions documented with parameters and returns and exposed in the API by @LegrandNico in #394
• Bump version from 0.2.12 to 0.3.0 by @LegrandNico in #395

Full Changelog: v0.2.12...v0.3.0