numerical/MC integration

[jonas-eschle]

Mostly for lower dimensions, we should have more than simple MC integration, as we have now.

Since things are different in TensorFlow and not every integration scheme can be implemented so easily, we have to experiment around with integrations techniques.

Currently in ConvolutionPDF, experimental:

• mc integration with map_fn for every element

Numerical schemes:

• gauss hermite quadrature from TFP

Adaptive MC schemes:

• there is ZMCintegration, which was supposed to be TF backend

Let's collect the findings of integrations with TF here.
@olantwin @Lakshan-Ram