Swap out `patsy` for `formulae`

[drbenvincent]

At the moment, the model formula provided to experiments is processed with patsy. This is no longer under active development.

We should switch to using formulae as a future proofing step. It also allows for hierarchical / random effect type models, should we see fit to use those in the future.

[drbenvincent]

Updating scope based on what we've learned since opening this: a straight engine swap turned out to be the wrong unit of work.

1. The hierarchical work (Add HierarchicalLinearRegression for Hierarchical DiD #860, adding hierarchical its #833) needs (x | g) mixed-effect syntax that patsy can't parse, so formulaic adoption is now pulled by features rather than preference alone.

2. Swapping engines experiment-by-experiment duplicates effort — each experiment stashes design internals privately, so an engine change touches all of them. Swap out patsy for formulaic #463's history demonstrates this directly.

New plan: settle the engine strategy inside the DesignTransform abstraction (#1094). The protocol hides the engine behind transform_x / transform_y, so patsy stays the default implementation today and a formulaic implementation becomes additive instead of a big-bang migration. #1094 is temporarily on hold while the hierarchical experiments land and tell us what the protocol must support; this issue stays open until that settles.