Replace term tuples with term vectors to reduce compilation latency

[matthieugomez]

Fitting a model with a never-seen formula currently costs ~0.1–0.8s of compilation, even when every term type is already precompiled. The reason is that formulas are encoded as tuples, so apply_schema, modelcols, coefnames, etc. get a fresh specialization for every distinct number and order of terms, and likewise every method taking a MatrixTerm{Ts<:TupleTerm}.

This PR represents term collections as Vector{AbstractTerm} instead of tuples of terms (or a single term when there was only one variable in the formula). This is a breaking change (0.8.0).
Changes

• + on terms returns Vector{AbstractTerm}. ~ (and @formula) always makes the rhs a vector, even for a single term (@formula(y ~ x).rhs == [term(:x)]), removing the lone-term special case. The lhs stays a bare term, and after apply_schema the rhs still collapses to a bare MatrixTerm, so y, X = modelcols(f, data) is unchanged.
• InteractionTerm and MatrixTerm are no longer parametric and store a Vector{AbstractTerm}. collect_matrix_terms returns a vector in the mixed matrix/non-matrix case.
• apply_schema over a vector loops term by term, with the same left-to-right FullRank semantics as the tuple broadcast.
• TupleTerm is removed; TermOrTerms is now Union{AbstractTerm, AbstractVector{<:AbstractTerm}}.
• Content-based == and hash for FormulaTerm, InteractionTerm, and MatrixTerm, since vector fields break the default === fallback. This also fixes formula equality after apply_schema (a former @test_broken).
• Version bumped to 0.8.0.

Latency

Timings below are for the first modelmatrix call on a formula shape not seen before, in a warm session on Julia 1.12, after warming up the pipeline 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))
# each line is a formula shape never seen before
@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);

StatsModels 0.7.10: every new shape recompiles the pipeline (~99.9% compilation time):

0.062946 seconds (272.38 k allocations: 13.183 MiB, 99.71% compilation time)
0.101134 seconds (621.89 k allocations: 30.417 MiB, 99.77% compilation time)
0.071856 seconds (345.43 k allocations: 16.959 MiB, 99.60% compilation time)
0.753103 seconds (3.40 M allocations: 168.672 MiB, 99.92% compilation time)

This PR:

0.000062 seconds (134 allocations: 53.391 KiB)
0.000066 seconds (175 allocations: 70.906 KiB)
0.013173 seconds (75.77 k allocations: 3.913 MiB, 98.90% compilation time)
0.374247 seconds (1.59 M allocations: 79.114 MiB, 99.89% compilation time)

Downstream impact

I ran the test suites of the main reverse dependencies against this branch.

• Pass unchanged: GLM, StandardizedPredictors.
• Fail to load because TupleTerm is gone (fix is a one-line signature swap to AbstractVector{<:AbstractTerm} per method): RegressionFormulae, MixedModels, RCall, AnovaBase, HurdleDMR, MixedModelsSerialization.
• Runtime failures from code that special-cases a tuple rhs (or isa MatrixTerm{Tuple{...}} checks), each a few lines: Econometrics, FixedEffectModels, ShiftShares, Metida, SurvivalAnalysis.

The only substantive migration is in MixedModels: the InteractionTerm{<:NTuple{N,CategoricalTerm}} dispatch (randomeffectsterm.jl:127,210) has to become a runtime check (all(t -> t isa CategoricalTerm, it.terms)), since interaction element types no longer live in the type domain.