Do not approve draft pull requests

[jaydev-CAstra]

Fixes #305

Problem

tag_approved only looks at the reviews: a draft PR with enough approvals gets the approved label, and when an approved PR is converted back to draft nothing removes the approved and ready to merge labels (no webhook handles that event, and the task would keep them anyway).

Reproduced with unit tests on a mocked PR (draft=True, two approvals): the task adds approved and schedules tag_ready_to_merge.

Fix

• tag_approved: a draft PR is never approved; if it carries the approved / ready to merge labels they are removed, same as when changes are requested.
• New webhook on_pr_draft on pull_request actions converted_to_draft and ready_for_review, which re-evaluates tag_approved for the PR.
• tag_ready_to_merge: the search query gets draft:false, so a draft PR that still has the approved label cannot become ready to merge.
• README and a 305.bugfix newsfragment.

Tests

• tests/test_tag_approved.py (new): approved label added, labels removed on changes requested, draft PR not approved, draft PR loses its labels.
• tests/test_on_pr_draft.py (new): the webhook schedules tag_approved.
• First commit, test infrastructure: set_config() restored the new value instead of the saved one, and test_switchable_method left BOT_TASKS modified, so any switchable task called by a later test was silently disabled (that is how the new tests first failed in the full run while passing alone). Both are now restored properly.

78 tests pass locally (the 3 test_build_wheels tests are skipped here because this machine's Python has no ensurepip, unrelated).

[MiquelRForgeFlow]

@legalsylvain It would be nice if you comment the reason why you close the PRs when closing them.

[legalsylvain]

@legalsylvain It would be nice if you comment the reason why you close the PRs when closing them.

I don't see the point in talking to a robot. And I don't want to waste my time doing that. Thanks.

[MiquelRForgeFlow]

I don't see the point in talking to a robot. And I don't want to waste my time doing that. Thanks.

I want to believe it's a real person that uses some LLM, not a real robot, so I think being a bit nicer to a contributor cannot be that difficult. @sbidoul, the PR fixes an issue you created, so you may want to say something here.

[legalsylvain]

I want to believe it's a real person that uses some LLM, not a real robot, so I think being a bit nicer to a contributor cannot be that difficult

I'll not make effort to be "nice" with such robotized contribution. Feel free to reopen.