Improve Gitlab::Database::Partitioning tests

Improve Gitlab::Database::Partitioning tests not to call private methods (using __send__):

• setup state using Gitlab::Database::Partitioning.registered_models
• move the responsibility to ensure that the registered models have partitioning strategy to register_models.
• and make sure we do not leak state outside tests

This brought up during the review of !72697 (merged):

• @furkanayhan started a discussion: (+2 comments)

  I'm confused with this? 🤔 We assign described_class.registered_for_sync to a variable, then stub the same method to the same variable, then use the same variable 😅

  I assume it's something to have the same return value from registered_models, but... should we do this? 🤔

• @krasio started a discussion:

  Similar to above, we can have this test like

        it 'reports metrics for each registered model' do
          expect_next_instance_of(partition_monitoring_class) do |partition_monitor|
            expect(partition_monitor).to receive(:report_metrics_for_model).with(models.first)
            expect(partition_monitor).to receive(:report_metrics_for_model).with(models.last)
          end  
  
          described_class.instance_variable_set('@registered_models', Set.new)
          described_class.register_models(models)
  
          described_class.report_metrics
        end  
      end

  though having to use instance_variable_set is not ideal, and means we miss method like reset or registered_models=.

• @krasio started a discussion:

  Does having to use __send__ mean we need to keep registered_models to public?

  BTW I think "ensure that the registered models have partitioning strategy" should be responsibility of register_models/register_tables - we should check and raise if the model provided does not respond to partitioning_strategy.