MERGE3 blocks unnecessarily in discovery when non-multicast transport is used

When MERGE3 uses TCP, it cannot multicast its INFO message, and therefore uses the discovery protocol (e.g. MPING) to fetch the targets to send the INFO message to.

Since we don't know how many responses to expect, we simply block for (min_interval + max_interval /2) ms. This is bad, as it delays the sending of INFO messages, which results in a partial merge as we're likely not to get responses from all members. This delays a full merge, e.g. when we have many singleton subclusters. A heavily split cluster will therefore likely require more merge rounds than necessary when using TCP, compared to (e.g.) UDP.

Solution:

• The discovery process should be reactive rather than blocking: instead of waiting for N seconds, we simply pass a function to the discovery protocol that gets invoked whenever a response has been received
• When the function gets invoked, it sends an INFO to the respective member
• This prevents 1 thread from blocking for N seconds

See [1] for details.
[1] https://github.com/belaban/JGroups/pull/389