Faster indexing with Categorical axes

With named columns, indexing based on a symbol slows down tight loops. DataFrames has the same problem. It's tough to get fast, type-stable code for this. Here is an example:

using AxisArrays

N = 10_000_000
A = AxisArray(rand(10000000,2), (1:N,[:a,:b]));

function f(x, j)
    res = 0.0
    for i in 1:size(x,1)
        res += x[i,j]
    end
    res
end

@time f(A, 2)  #  elapsed time: 0.028395385 seconds (117 kB allocated)

@time f(A, :b) #  elapsed time: 0.147722792 seconds (364 kB allocated)

I tried using types as column names to speed things up, but the resulting code isn't type stable, and it's even slower. Here's my attempt:

## Bunch of kludgy stuff
AxisArrays.checkaxis{T}(::Type{Val{T}}) = true
stagedfunction AxisArrays.axisindexes{T,S}(::Type{Val{T}}, ::Type{Val{S}})
    i = findfirst(T, S)
    :($i)
end
Base.length{T}(::Type{Val{T}}) = length(T)
Base.size{T}(::Type{Val{T}}) = size(T)
Base.size{T}(::Type{Val{T}},i) = size(T,i)

B = AxisArray{Float64,2,Array{Float64,2},(:row,:col),(UnitRange{Int64},DataType)}(rand(N,2), (1:N,Val{(:a,:b)}));

@time f(B, Val{:b}) # elapsed time: 0.982083383 seconds (461 MB allocated, 3.26% gc time in 20 pauses with 0 full sweep)
@time f(B, 2)       # elapsed time: 0.019993164 seconds (88 kB allocated)

Start by reproducing the issue's f(A, 2) and f(A, :b) benchmarks with the supplied AxisArrays example, then inspect the indexing path for named axes. Done means symbol-based indexing is faster and type-stable without regressing integer indexing; the issue names no files or tests to run.