[DOC] Make Joiner example 07 more compelling

[Vincent-Maladiere]

Describe the issue linked to the documentation

As @jeromedockes mentioned in #1145, example 07 has several flaws:

In a second time, we can work on improving that example and maybe using a different dataset. Indeed, at the moment
- it downloads a very large dataset
- the join with weather data does not improve predictions significantly IIRC
- the predictions are not super far from chance level which makes the example less compelling IMO

Suggest a potential alternative/fix

We could try to either:

• Find another dataset on which fuzzy joining would boost performance
• Change the task to joining, without focusing on learning. A common use case for people working with databases is to create "proxy keys" from multiple columns. This technique is helpful to join tables coming from different storing systems or databases, where there isn't a foreign key to join on. Without necessarily having a downstream learning task, we can quantify the fuzzy join with e.g. recall and precision, by defining "false positive" as incorrect joins and "false negative" as missed joins.