Madminer

Bug fixes
- Fixed a bug in the setup.py introduced in v0.7.3, see 432

Internal changes:
- Disentangled and cleaned up the dependencies. `pip install madminer` will now only install the dependencies that are necessary for the core library, not the additional dependencies used in the examples or documentation. New options like `pip install madminer[examples]` can be used to install these.

API changes:
- Improved the support for Python 2 (a necessary evil for MadGraph): in the `MadMiner.run()` functions, using `python2_override=True` without providing the keyword `initial_command` will now lead to the Python 2 executable being prepended to the Python path. See 423.

Documentation:
- Updated docstrings, see 420.

We're very happy to welcome our new contributor Sinclert, who initiated and implemented both of these fixes!

Bug fixes:
- Various small bug fixes.

Internal changes:
- Refactoring the code structure, moving to more submodules. The API and functionality is unchanged. One can now directly import from the `madminer` namespace, e.g. as `from madminer import MadMiner` (but the old `from madminer.core import MadMiner` will still work).

New features:
- New `madminer.likelihood` class will let the user define more powerful likelihood functions, construct a profile likelihood, and fit with more flexibility.

Documentation:
- Updated tutorial notebooks.

New features:
- New Fisher information geometry functionality in the `madminer.fisherinformation.InformationGeometry` class.
- The Fisher information can now be computed from likelihood ratio estimators as well.
- Added `early_stopping_patience` keyword to ML training functions.

Bug fixes:
- Fixed a bug in `DelphesReader`.

Documentation:
- Updated and fixed tutorials.