Decoupling the change detection and alerting logic

[MohamedBilelBesbes]

This PR is dedicated to separate the logics of the change point detection and the alerting.

• Separated the logic of change point detection from the logic of alert creation: run_detection handles pure detection with no database side effects, and create_alerts_from_results is the sole function responsible for persisting alerts, making it possible to query the detection pipeline without triggering any alerting (meaning that Sherlock could use it).
• Made run_detection configurable to operate in two modes: a voting mode (voting_strategy="equal" or "priority") that runs the full ensemble and applies a consensus strategy, and a single-method mode (voting_strategy=None) that returns change points flagged directly by one named CPD method without involving the voting system.
• Moved the responsibility of selecting which methods participate in the voting system from build_cpd_methods into generate_new_test_alerts_in_series via the voting_methods param.
• Updated the existing deduplication guard unit test to match the new decoupling logic, where build_revision_data and run_detection are now called explicitly and separately instead of through the previously coupled vote function.
• Added a new unit test to verify the single-method mode of run_detection, using Mann Whitney U as the detector and asserting that results come exclusively from that method's change_detected flags with no voting involvement.