[Task] Improve CI workflows #3517

Some issues that we can address to improve CI:

1. CI is currently quite overloaded for new PRs. Can we reconsider the specific selection of tests that will be run for general PRs? This could involve improving the filters in the intelligent testing workflow, or simply reducing the matrix of tests.
2. The canary lane is now somewhat separated from the rest of intelligent testing. We should consider integrating this better.
3. Should we look at the functional tests and if they are still representative and covering our standard pipeline? or can we improve?
4. Are all tensorflow (smoke) tests still needed or can we reduce these?