Support for refitting LA with option to override or update

[aleximmer]

Changes:

• make mean a quantity that falls back to prior when not yet fit and is set on .fit() so that .log_prob() is correct
• add log_prob(value) to compute log density under LA for regularization. This falls back to prior before fitting the LA
• add .fit(train_loader, override=True) where override means to reset Hessian approximation while override == False leads to adding up curvature, e.g., for continual learning

[runame]

Looks great, this really simplifies continual learning! I wrote a few minor comments.

Optional, probably not needed:

• Test repeated fit for last-layer Laplace
• Test initialization of mean and precision to prior