Refactor tasks: GPU/Dask/Ray/Server support

[maartenbreddels]

This PR completely refactors how task work. Now a task (e.g. an aggregation) is defined, and can be cheaply serialized, and can be executed in several ways. This PR focusses mostly on the server-side and CPU execution, but local experimental branches have shown this to work well with Dask or Ray for distributed computing, although this requires #548 for efficient serialization of data/dataframes, and some refactoring of groupby/join/take etc. It also opens to door to other task executions, e.g. on the GPU, for example using the libraries powering cuDF.

[maartenbreddels]

This failure is so odd:
https://travis-ci.org/vaexio/vaex/jobs/640451568?utm_medium=notification&utm_source=github_status

Sometimes we get 6, sometimes 7 columns.

[maartenbreddels]

Failure was due to a cached file in ~/.vaex/data/
windows failure keeps coming back @JovanVeljanoski
https://ci.appveyor.com/project/maartenbreddels/vaex-4gh4b/builds/30636573/job/fhpi3r4hnt3lhmmm

    @pytest.mark.skipif(((1,17,0) <= version <= (1,17,5)) and platform.system().lower() == 'windows', reason="strange ref count issue with numpy")
    @pytest.mark.skipif(((1,17,0) <= version <= (1,17,3)) and platform.system().lower() == 'linux' and sys.version_info[:2] == (3,6), reason="strange ref count issue with numpy")
    def test_robust_scaler():
        x = np.array([-2.65395789, -7.97116295, -4.76729177, -0.76885033, -6.45609635])
        y = np.array([-8.9480332, -4.81582449, -3.73537263, -3.46051912,  1.35137275])
        z = np.array([-0.47827432, -2.26208059, -3.75151683, -1.90862151, -1.87541903])
        w = np.zeros_like(x)
    
        ds = vaex.from_arrays(x=x, y=y, z=z, w=w)
        df = ds.to_pandas_df()
    
        features = ['x', 'y']
    
        scaler_skl = RobustScaler()
        result_skl = scaler_skl.fit_transform(df[features])
        scaler_vaex = vaex.ml.RobustScaler(features=features)
        result_vaex = scaler_vaex.fit_transform(ds)
    
>       np.testing.assert_array_almost_equal(scaler_vaex.center_, scaler_skl.center_, decimal=0.2)

[JovanVeljanoski]

Ahh.. so the windows CI starts using numpy 1.18.1, and this issue is still there.. :(

Linux also has numpy 1.18.1 but there it passes..