Add explicit classes for structured and unstructured formula variants.

[matthewwardrop]

This patch adds explicit classes for the structured and unstructured formula variants. I was on the fence about pulling the trigger on this, but the issues identified in #213 make the increase in local complexity worth it.

This is a largely backward compatible change, unless code is hard-coded to expect structured instance methods like ._map and _has_structure. I doubt anyone is using them, but just in case, before merging I will add in backward compatible shims for these methods that raise the appropriate deprecation warnings; and ultimately we'll remove these shims in version 2.0 (whenever that happens). Users need never intentionally interact with the new SimpleFormula and StructuredFormula classes.

As a result of these changes, code like:

Formula(lhs=Formula(["y"]), rhs=Formula("1 + z:x + x + y + g:h", _ordering="none"))

will retain the ordering of the nested formulae rather than reordering them; due to the new distinction between List[Term] (which is still reordered according to the local ordering policy) and SimpleFormula which is responsible for its own ordering.

TODO:

• Deprecated shims for Structured methods and attributes on the new SimpleFormula class.
• Unit tests
• Documentation

closes: #213

[matthewwardrop]

@bashtage I'll clean this up in over the next few days, but let me know if this breaks anything in statsmodels. Even if it doesn't, I plan to add the backwards compatible shims with deprecation warnings (one of which is already present: .root).

[bashtage]

Thanks for getting this in. I just need to finalize cubic splines and then I think statsmodels will be able to pass tests against main. It already is passing on main + cubic + ordering. I'll let you know if there are any issues after I pull this one in instead of my ordering patch.