Better contrast specification

Two possible improvements in how we work with contrasts:

• (minor) Maybe we should have a way to specify a contrast symbolically, like linear_constraint but just the linear part, not the = constant part.
• (major) There should be some way to specify contrasts and constraints in terms of predictions (which are invariant wrt coding), rather than predictors (betas). Like it should be possible to say "the difference between an item with a=1 and an item with a=2", and the model will spit out a matrix encoding this -- or, crucially, if the model has an interaction between a and b, then it will say "that's not estimable, you can't leave b unspecified". Possibly we also want to be able to say "tell me the derivative of the prediction wrt x1", since that's also one of the things that betas encode.

Syntax for this will be tricky. Possibly just an array of weights + a list of data dicts?

I need to brush up on the estimable contrast literature...

Start by reviewing the existing linear_constraint entry point and the contrast-related API, then read the estimable contrast literature referenced in the issue. Before implementation, resolve whether the scope includes symbolic linear parts, prediction-based contrasts, interaction estimability checks, and prediction derivatives; done means an agreed syntax and behavior for the selected cases.