Reduce overhead of loading class definitions

[guusdk]

When sharing data over the cluster, each serialization request seems to load the Class definition of the associated data. This can quickly add up.

On a fairly busy cluster, I noticed that pretty much all contention around Monitors/Locks related to Hazelcast waiting for classes to be loaded. The screenshots taken below show that this is by far the most resource intensive operation, and that the amount of CPU cycles spent on loading classes almost equals the amount of cycles spent on actually reading the data!