Improve pydelfi implementation and examples

[maho3]

I gave a tutorial of ltu-ili, in which several folks tried the pydelfi implementation. It pointed out a number of ways in which the implementation could be improved, made more similar to the sbi and lampe backends, and general things to make its' use clearer.

Things that have changed (feel free to use this as a checklist for review):

• [ ] DelfiRunner training summaries are now structured in the same way as SBIRunner
• [ ] np.array casting all the tensorflow distribution params, so they can be specified with floats
• [ ] Better handling of DelfiRunner.out_dir when it is set to None
• [ ] Fixing the shape of Delfi.potential for the jupyter notebook
• [ ] Adding a pydelfi jupyter notebook for making things clear to the users

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 96.85%. Comparing base (79d6f8a) to head (c6b0e1c).

:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #145 +/- ## ========================================== + Coverage 96.82% 96.85% +0.02% ========================================== Files 23 23 Lines 1605 1620 +15 ========================================== + Hits 1554 1569 +15 Misses 51 51 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.