Separate the change of sampling space (θ) from VarInfo

[xukai92]

The target of this proposal is to reduce the memory used for storing the theta history as less as possible. Let's say θ is the variables which are associated with the sampler.

Key points:

1. Sampler only changes and keep a track of the θ but not VarInfo.

• VarInfo is only treated as a layer that sampler uses to interact with model.

2. Each time sampler want to apply θ to the model, it firstly pass θ to VarInfo, and run the model with VarInfo and sampler.

• It's sampler's responsibility to use the correct θ (e.g. use an old θ if a previous step is rejected) when evaluating the model.

3. When passing VarInfo to another sampler, all the sampler need to do is vi[spl] = θ and pass this vi to the other sampler.

• During the process of one sampler, it will never change values belong to other samplers in this case because vi[spl] = θ will not interact other space.

Running model is like

evalute(θ, model, spl, vi) = begin
  vi[spl] = θ
  model(vi=vi, sampler=spl)
end

Layer concept is like

---------------------------------
|              Model            | ----------> Sample
---------------------------------
         ↑                ↑              
-------------------       ↑
|     VarInfo     |       ↑
-------------------       ↑
         ↑                ↑
-------------------       ↑
|     Sampler     | <------
-------------------

[xukai92]

Done in #254