Refactor markov

[sglyon]

This implements a few things:

• reorganize markov chain tools into the markov folder
• loosen some type constraints to be AbstractArray instead of only Array
• Add support for sparse transition matrices in the MarkovChain type.

[sglyon]

This is the one main thing I don't feel great about in this PR. Notice that I added full around each row of the stochastic matrix.

DiscreteRV uses a cumsum over an AbstractVector to give us the transition to the next state. In Julia 0.4 there is no SparseVector type (only sparse matrix), so to get this to be an AbstractVector we needed to make it dense before passing to DiscreteRV. The thing that makes it ok is that cumsum is an operation that will make any sparse array dense, after the first non-zero element. So we probably are more efficient working directly with the dense array than we would have been working with a dense sparse array, but I just wanted to point this out.

Also, I checked the native machine code for this function with a dense mc.p before and after making this change and the machine code was identical. this means that when mc.p is already dense the jit compiler completely removes the no-op associated with calling full and we have 0 performance hit.

[sglyon]

Note, travis failures here are only on julia 0.5 (downstream issue with a dependency not loading on unreleased julia).

I have updated the travis.yml file in #100 to allow failures on 0.5, but haven't duplicated that commit here. If there are no objections I'm ok to merge this.

[sglyon]

Added a few more things:

• state_values as a field on MarkovChain. This is any AbstractVector and defaults to 1:size(p, 1), where p is the transition matrix
• Methods simulate_values to do a simulation and report state_values instead of state indices
• simulate now always returns a Matrix (as does simulate_values)
• To do a single simulation and get a vector use simulation or value_simulation
• Extended DiscreteRV to allow for q and Q to be different types of AbstractVector. I did this so q could be a SubArray in simulate! and simulate_values! to avoid allocations
• Added inner constructor check on DiscreteRV that sum(q) == Q[end] == 1.0

[sglyon]

failing tests are a result of a dependency not working on the dev version of Julia. In another PR I have specified that these tests are allowed to fail. We can merge this without waiting for tests to pass.

[sglyon]

Last change for now, I promise: tauchen and rouwenhorst now return instances of MarkovChain instead of a state_values, p tuple

[oyamad]

This doesn't have to belong to this PR, but random_markov_chain and random_stochastic_matrix in random_mc.jl should support sparse matrix (by the way, this file should go to markov/ too). Compare the python version.

[sglyon]

Good point, see #102