Index eltypes unioned on empty `KeyedArray`

There seems to be a weird bug where the axiskey eltypes get unionized when converting a KeyedArray with an empty dimension to a table. Perhaps we should define a Tables.schema method for a KeyedArray to avoid this problem?

julia> ka = KeyedArray(rand(0, 3); x=1:0, y=[:a, :b, :c])
2-dimensional KeyedArray(NamedDimsArray(...)) with keys:
↓   x ∈ 0-element UnitRange{Int64}
→   y ∈ 3-element Vector{Symbol}
And data, 0×3 Matrix{Float64}

julia> Tables.schema(DataFrame(ka))
Tables.Schema:
 :x      Union{Int64, Symbol}
 :y      Union{Int64, Symbol}
 :value  Float64
 
 julia> ka = KeyedArray(rand(0, 3, 2); x=1:0, y=[:a, :b, :c], z=["foo", "bar"])
3-dimensional KeyedArray(NamedDimsArray(...)) with keys:
↓   x ∈ 0-element UnitRange{Int64}
→   y ∈ 3-element Vector{Symbol}
□   z ∈ 2-element Vector{String}
And data, 0×3×2 Array{Float64, 3}

julia> Tables.schema(DataFrame(ka))
Tables.Schema:
 :x      Union{Int64, String, Symbol}
 :y      Union{Int64, String, Symbol}
 :z      Union{Int64, String, Symbol}
 :value  Float64
 
 julia> eltype.(axiskeys(ka))
(Int64, Symbol, String)

Start by running the KeyedArray examples in the issue and inspect how KeyedArray values are converted to tables, including the Tables.schema interface. Check the axis key eltypes for empty dimensions and compare them with the schema produced by DataFrame(ka). Done means the schema preserves each axis key type instead of unioning unrelated key types.