Add entropy regularised policy distillation

[pwuethri]

• [x] add code for entropy regularised policy distillation
• [ ] bring script to run in machina
• [ ] test if code works using pendulum v0
• [ ] apply autopep8 to all changed files

[pwuethri]

@rarilurelo Should i put all the policy distillation approaches into the same example file? We can make a user choose which approach he/she wants to use and then calculate the loss accordingly in the algos/policy_distillation.py file.

I worry that otherwise, the example and algos directory will become cluttered with policy distillation files.

But if you prefere, we can keep each distillation approach separated by different scripts. Which way do you prefere?

[rarilurelo]

I don't care about cluttered files for now. Finishing implementation of distillation, you can try to put them together.

[pwuethri]

Understood

[pwuethri]

Grads were calculated in a wrong manner -> needs to be fixed

[pwuethri]

The calculated gradients are too big (10e3 - 10e4), leading to huge parameter update.

[pwuethri]

total reward term is approx 400'000...

[pwuethri]

student llh seems to be reasonable