In `ReferenceProcessor`, the multi-thread degree of ref-processing is subject to multiple constraints:

1. `mt_processing_degree` passed to the constructor
2. the global boolean flag `ParallelRefProcEnabled`
3. #active workers in the worker-pool from each collector

The effect of 1 & 3 is manifests via `_num_queues` inside the constructor and `set_active_mt_degree` method. Hence the body of the following method:

```
bool ReferenceProcessor::processing_is_mt() const {
  return ParallelRefProcEnabled && _num_queues > 1;
}
```

Such arrange leads to wasted memory in the situation where `ParallelRefProcEnabled == false` && `ParallelGCThreads > 1` && `ParallelGCThreads > ConcGCThreads` -- unnecessary ref queues are created and those extra queues are never actually used.

One way to remedy the situation is to include the effect of 2 in `_num_queues` inside the constructor and drop the API mutating `_num_queues` (`set_active_mt_degree`). In other words, the ref-processor would have access to the worker-pool from the collector and take constraint 3 into account in each subphase of ref-processing while adjusting `_num_queues`. Consequently, `_num_queues` will appear as external immutable.

Additionally, `processing_is_mt()` becomes `(_num_queues > 1)` essentially.