Added missing values handling.

[trappmartin]

This PR adds missing values handling for the internal Turing.Chain object and Sample object. This work is required for BNP related models.

Changes:

• Removed value2 field which contains redundant information.
• Changed getindex function to use MCMCChain.names2inds
• Chain is now immutable (I hope this doesn't break anything!)
• value field supports missing values.
• Refactored constructor and flatten functions.
• Added more tests.

Supported indexing:

@model testmodel() = begin
    x ~ Normal()
    y ~ Dirichlet([0.5, 0.5])
    z ~ Wishart(7, [1 0.5; 0.5 1])
end

...
chain = sample(mf, SMC(100))

chain["x"] # => 100 x 1 x 1 matrix
chain[:x] #  == chain["x"]
chain["y[1]"] #  => 100 x 1 x 1 matrix
chain[["y[1]", "y[2]"]] #  => 100 x 2 x 1 matrix
chain[:y] # == chain[["y[1]", "y[2]"]]
chain["z[1, 1]"] # => 100 x 1 x 1 matrix
chain[:z] # => 100 x 4 x 1 matrix
reshape(mean(chain[:z], dims = [1]), 2, 2) # average z matrix

[trappmartin]

@cpfiffer This PR is ready to merge from my side. Could you have a look? Thx!

[cpfiffer]

Looks good to me. I think we can merge, unless anyone else has anything to say on this.

[yebai]

Nice, thanks @trappmartin!