Type instability for `vec[1, end]`

[Ickaser]

Describe the bug 🐞

Accessing a recursive array with [1, end] is type-unstable. Was type stable on v3.41 and earlier, broken on 3.42 and later. Related to @JoshuaLampert 's recent work, it seems.

Expected behavior

Type stability for array accesses.

Minimal Reproducible Example 👇

vecvec = VectorOfArray([fill(2.0, 2) for i in 1:10])
access_end(vec) = vec[1, end]
@code_warntype access_end(vecvec)

yields

MethodInstance for access_end(::VectorOfArray{Float64, 2, Vector{Vector{Float64}}})
  from access_end(vec) @ Main REPL[4]:1
Arguments
  #self#::Core.Const(Main.access_end)
  vec::VectorOfArray{Float64, 2, Vector{Vector{Float64}}}
Body::Any
1 ─ %1 = Base.lastindex(vec, 2)::Core.PartialStruct(RecursiveArrayTools.RaggedEnd, Any[Core.Const(0), Int64])
│   %2 = Base.getindex(vec, 1, %1)::Any
└──      return %2

Error & Stacktrace ⚠️

Digging around with Cthulhu, I see:

that is, starting at https://github.com/SciML/RecursiveArrayTools.jl/blob/6519ecf4fbf69a6134051d1fa11ad5685dd93c3a/src/vector_of_array.jl#L838, `cols`, `prefix`, `resolved`, `inner_nd`, `padded` all are getting boxed somehow and killing the inference.