`@where` does not preserve CategoricalArray

Julia v0.6, DataFrames, CategoricalArray and Query masters

julia> using DataFrames, Query

julia> df = DataFrame(a = categorical(["a", "b", "a", "c"]))
4×1 DataFrames.DataFrame
│ Row │ a │
├─────┼───┤
│ 1   │ a │
│ 2   │ b │
│ 3   │ a │
│ 4   │ c │

julia> typeof(df[:a])
CategoricalArrays.CategoricalArray{String,1,UInt32,String,Union{}}

julia> dff = df |> @where(_.a != "b") |> DataFrame
3×1 DataFrames.DataFrame
│ Row │ a │
├─────┼───┤
│ 1   │ a │
│ 2   │ a │
│ 3   │ c │

julia> typeof(dff[:a])
Array{CategoricalArrays.CategoricalValue{String,UInt32},1}

IIUC Array{CategoricalValue} is suboptimal in terms of storage. Also CategoricalArrays API like levels(dff[:a]) is not available.

Reproduce the issue with the Julia commands in the report, using DataFrames, CategoricalArray, and Query. Trace the @where pipeline to find where the categorical column changes representation; done means the filtered result retains a CategoricalArray and supports levels(dff[:a]) as expected.