Port to CategoricalArrays and Missings

[nalimilan]

Breaking change: rename usena to skipmissing, and change default
to keep missing values. Also rework CategoricalArray method to make
it simpler. Ensure that type inference works when input arrays do not
support missing values. Improve tests.

@quinnj Working with a DataFrame importing using CSV.jl gives corrupted WeakRefString objects in the following example:

using DataFrames, CSV
iris = CSV.read(joinpath(Pkg.dir("DataFrames"), "test/data/iris.csv"), categorical=false);
iris[:LongSepal] = iris[:SepalLength] .> 5.0
tab = freqtable(iris, :Species, :LongSepal)

Sometimes, printing tab fails since the dimension names appear to have been freed. That's really problematic. Any plans to prevent this?

EDIT: fixed by JuliaData/WeakRefStrings.jl#17 (comment).

[coveralls]

Coverage decreased (-1.9%) to 95.455% when pulling cfdb1c8 on nl/dataframes0.11 into fca6888 on master.

[coveralls]

Coverage decreased (-1.9%) to 95.455% when pulling cfdb1c8 on nl/dataframes0.11 into fca6888 on master.

[coveralls]

Coverage decreased (-1.9%) to 95.455% when pulling fa62b1b on nl/dataframes0.11 into fca6888 on master.