fix(consumer): loopCheckPartitionNumbers may not cause rebalance after adding new partitions

[napallday]

fix #2460

[napallday]

I know there is a data race in the newly-added test file

func (h *loopCheckerHandler) Setup(s ConsumerGroupSession) error {
	if len(s.Claims()["my-topic"]) == 1 {
		if err := h.group.client.RefreshMetadata("my-topic"); err != nil {
			h.Error(err)
		}
		// change RefreshFrequency so that loopCheckPartitionNumbers won't return directly
***		h.group.client.(*nopCloserClient).Client.(*client).conf.Metadata.RefreshFrequency = 5 * time.Millisecond
		go h.group.loopCheckPartitionNumbers([]string{"my-topic"}, s.(*consumerGroupSession))
	} else {
		h.cancel()
	}
	return nil
}

The added test will fail in the old versions without a race detector, but I haven't managed to write a test to reproduce the issue without a data race.

If it's difficult to resolve the data race, how about we just remove the unit test if the changes look good to you guys? @dnwe @hindessm

[napallday]

I found another critical bug in our company, which is related to this issue.
I'll open another MR to fix both of them.