[review] Dask performance tuning + Dask plugin + Smart Cache handling + Composite Node performance tuning + Template Node

[yidong72]

In this PR, I made following changes:

1. By default, the Dask dataframe meta data is using the meta_setup.
2. Added the necessary persist node. The changes of 1-2 make existing dask workflow 2 times faster.
3. Use the ml.xgboos for XGBoost strategies.
4. Added the SimpleParallelNode to make parallel computation easy.
5. Added a notebook as an example use of SimpleParallelNode
6. Improves the DaskComputeNode to handle multiple ports simultaneously.
7. Move the DaskComputeNode, SimpleParallelNode and PersistNode into a separate Greenflow plugin
8. Change the core plots to matplotlib as bqplot is not stable enough.

I tested all the notebooks.

[yidong72]

I the latest change, I added SimpleNodeMixinNode to simplify the node code and added caching mechanism to improve the performance. I tested all the notebooks. It is working and ready for review!

[yidong72]

This should be my last round of change.

The composite node only build the graph once in the update function so ports_setup, meta_setup, conf_schema can be called cheaply.

GridRandomSearchNode is also fixed to removed the strong requirement that all upstream graphs need to be serializable.

[avolkov1]

I'm still reviewing and testing. Left two comments so far.

[avolkov1]

I left a few comments. The logic in the SimpleNodeMixin seems to me like it really should be part of NodeTaskGraphMixin. Let me know what you think.

Can you explain a bit more regarding the SimpleNodeMixin, especially its update, ports_setup and meta_setup methods, if you disagree with me regarding that belonging in the NodeTaskGraphMixin? I think all nodes should have this logic injected when they are within a task graph. It shouldn't be necessary to inherit explicitly from any mixin for a Node's implementation.

[yidong72]

I made a few changes to make the template state private. I checked all the notebooks and unit tests.

[avolkov1]

Looks good. Everything works. I'll merge this.