Mxfusion

The bigger changes are:
* 133 Changing variable broadcasting for factors
* 153 Simplify serialization to use a simple zip file

Other than that its mostly bug fixes and documentation changes.
* 144 Add details to inference and serialization documentation
* 148 Fix a bug for SVGP regression with minibatch traning
* 81 Reduce num_samples in uniform non-mock test to 1000
* 137 Changed transpose property to function.
* 135 Bug fix in function evaluation

- Add the tutorial for Gaussian process regression.
- Fix empty operator bug.
- Fix bug to allow the same variable for multiple inputs to a factor.
- Add module serialization.
- Fix the bug that causes the failure of documentation compilation.
- Fix the bug: the inference methods of GP modules do not handle samples.
- Update issue templates.
- Add license headers to all files.
- Add the getting started notebook.
- Remove the dependency on future.
- Update the Inference documentation.
- Implement Dirichlet distribution.
- Add logistic variable transformation.
- Implement Expectation inference.
- Fix the bugs related to dtype.
- Validate shape of array variables.
- Fix divide by zero error if max_iter < n_prints.
- Add multiply kernel.

* Improve the printing messages of gradient loop and bug fix for GluonFunctionEvaluation
* Add Gamma Distribution
* GP Modules enhancement
* Implement Module, GPRegression, SparseGPRegression and SVGPRegression…
* Update the README
* Add score function for variational inference enhancement
* Refactor the interface of inferece for module
* Add score function inference
* Implement common MXNet operators (dot product, diag, etc. in MXFusion enhancement
* Add support for MXNet operators.
* Cleanup kernel function and function wrappers for copying enhancement
* Implement the base class MXFusionFunction