suboptimal inference for `nextind`, `prevind` on `AnnotatedString`

[nsajko]

AnnotatedString is the only AbstractString subtype provided with an installation of Julia for which nextind, prevind have suboptimal inference. Script:

infer_return_type(f, S::Type) = Base.infer_return_type(f, Tuple{S, Vararg})

for S ∈ (Base.AnnotatedString, LazyString, String, SubString, SubstitutionString)
    @show S
    for f ∈ (nextind, prevind)
        print(' '^4)
        @show f
        print(' '^8)
        print("* ")
        @show infer_return_type(f, S)
    end
end

Output:

S = Base.AnnotatedString
    f = nextind
        * infer_return_type(f, S) = Any
    f = prevind
        * infer_return_type(f, S) = Any
S = LazyString
    f = nextind
        * infer_return_type(f, S) = Int64
    f = prevind
        * infer_return_type(f, S) = Int64
S = String
    f = nextind
        * infer_return_type(f, S) = Int64
    f = prevind
        * infer_return_type(f, S) = Int64
S = SubString
    f = nextind
        * infer_return_type(f, S) = Int64
    f = prevind
        * infer_return_type(f, S) = Int64
S = SubstitutionString
    f = nextind
        * infer_return_type(f, S) = Int64
    f = prevind
        * infer_return_type(f, S) = Int64

NB: the specific example shown above will be fixed when #57605 is fixed, however someone (maybe me, not sure) should probably look into the root cause in the AnnotatedString implementation.

[nsajko]

For completeness, this is caused by ncodeunits(::AnnotatedString).