Store terms in vectors to reduce compilation latency

[matthieugomez]

StatsModels currently recompiles much of the formula pipeline when the number or order of terms changes. Those differences change the tuple type, giving apply_schema, modelcols, and related methods new specializations.

This PR stores term collections in Vector{AbstractTerm}. It is a breaking change for 0.8.0; the @formula syntax stays the same.

Changes

• + builds a vector of terms. Formula rhs values are always vectors, including single-term formulas.
• InteractionTerm and MatrixTerm store vectors and are no longer parametric. collect_matrix_terms returns a vector when matrix and non-matrix terms are mixed.
• TupleTerm is removed. TermOrTerms accepts a term or an AbstractVector{<:AbstractTerm}.
• Schema application remains left to right. Empty term vectors are supported; empty interactions throw an error.
• Formula, interaction, and matrix terms compare and hash by content. FunctionTerm gets a hash consistent with its equality method.

Latency

These are first calls for new formula shapes in a warm Julia 1.12 session. Before measuring, I called modelmatrix with y ~ 1 + a and y ~ 1 + a + e.

d = (y=rand(1000), a=rand(1000), b=rand(1000), c=rand(1000),
     e=repeat(["u", "v"], 500))

@time modelmatrix(@formula(y ~ 1 + a + b), d);
@time modelmatrix(@formula(y ~ 1 + a + b + c), d);
@time modelmatrix(@formula(y ~ 1 + a + b + e), d);
@time modelmatrix(@formula(y ~ 1 + a + b + c + a&b), d);

Times in milliseconds, in the order above:

Call	0.7.10	This PR
1	62.95	0.06
2	101.13	0.07
3	71.86	13.17
4	753.10	374.25

Almost all of the time on 0.7.10 is compilation. The first two calls avoid it on this branch; the categorical and interaction examples still compile. Changing the number of variables in an interaction can still trigger compilation.

Downstream impact

I ran the reverse-dependency test suites against this branch:

• GLM and StandardizedPredictors pass unchanged.
• RegressionFormulae, MixedModels, RCall, AnovaBase, HurdleDMR, and MixedModelsSerialization fail to load because they reference TupleTerm. Those signatures need to use AbstractVector{<:AbstractTerm}.
• Econometrics, FixedEffectModels, ShiftShares, Metida, and SurvivalAnalysis fail at runtime where they expect a tuple rhs or a parametric MatrixTerm.

MixedModels also dispatches on InteractionTerm{<:NTuple{N,CategoricalTerm}} in randomeffectsterm.jl:127,210. It needs a runtime check such as all(t -> t isa CategoricalTerm, it.terms).

Docs CI currently fails during dependency resolution because GLM requires StatsModels 0.7. The doctests pass locally.