Reference Issues/PRs

What does this implement/fix? Explain your changes.

The 20newsgroups classification example is our only document classification
example and it can/should be more didactic. Therefore, this PR reworks it by:

• adding a loading function to easily remove text metadata when fetching dataset
• adding a demo on a particular classifier before benchmarking the rest of them:
  • compare feature effects
  • show pollution in the dataset introduced by the metadata / motivate preprocessing dataset
  • use ConfusionMatrixDisplay to visualize results
• adding text to provide interpretation of intermediate steps
• removing similarly performing classifiers to avoid distractions and overlap in plots
• removing the feature selection tools (they appear to have little to none effect)
• making the benchmark function accept customized names

Any other comments?

The demo uses RidgeClassifier as it has a relatively low training time, but any classifier with a coef_ attribute may work.

This example uses TfidfVectorizer only. A comparison of its performance with respect to HashingVectorizer will be craft in another example.