Several small fixes

[kbattocchi]

This PR makes a few small changes

• Reduce number of simulations in dowhy refutation to prevent notebook test timeouts
• Changes our sklearn-based linear models to remove all use of the normalize init arg (deprecated in sklearn 1.0, removed in sklearn 1.2)
• Cap protobuf versions to work around tensorflow compat issue ([Python] Release 4.21.0 broke multiple Google Cloud client libraries ("TypeError: Descriptors cannot not be created directly.") protocolbuffers/protobuf#10051)
• Removes Python 3.6 support (EOL was last December and several dependencies have already dropped support)
• Enables using Sphinx >= 5.0 for documentation generation
• Updates the dowhy documentation links to point to their new py-why address