We should ensure that when CompletionStages are completed they are continued only on Infinispan threads

When a user can supply a CompletionStage to delay operations this can cause a possibility of that being completed on a different thread than expected. The new non blocking listeners will especially be affected by this (ISPN-9614). This is also problematic for CacheLoader SPI and any future API that takes a Publisher as an argument from the user.

This issue should at least fix the upcoming non blocking listeners as discussed at https://github.com/infinispan/infinispan/pull/6401#issuecomment-475181484.