Remove remote dataset retrieval for tests

[Nacho888]

Remove all calls to fetch_california_housing from the tests, so that no internet connection is required for the execution of these. This is a small optimization but more in line with the best practices of avoiding making these type of connections while executing the test suite.

It removes an external dependency, and might make it slightly faster - at the expense of the addition of an extra file. This file will only be available for tests and the won't be bundled in the end-user package.

This pattern might need to be extended to other tests that make use of other remote datasets.

[solegalli]

Hi @Nacho888

Thank you very much for this PR. We definitely need to step away from making urlib calls in our tests.

I think for simpler transformers like the discretiser, we could replace the california housing by a hard coded dataframe, and then we don't store a csv file.

For more complex transformers, like MRMR, we could try creating an artificial dataset with make_classification and see if that resolves the issue, before committing to storing a csv file as part of the test suite.

It would be good as a start, to understand which tests rely of this dataset, so we know how big a change we need if we were to replace it completely.