AdaptivePoolingAllocator EventLoop Magazine's affinity

[franz1981]

Motivation:

AdaptivePoolingAllocator uses some shared Magazines hoping thread's id distribution to improve contended behaviour, but event loops threads can just uses their own Magazine(s) and save 2 atomic operations (write lock/unlock) in the hot path, further improving performance.

Modification:

Implements dedicated event loop Magazines

Result:

Better allocation performance for event loop threads

[franz1981]

This is trying hard to make close right(ish) - that's why it doesn't allow to just use FastThreadLocalThreads, but I can relax this a bit, accepting FastThreadLocalThreads which cleanup their FastThreadLocal, too: but clearly it means that closing the allocator won't do anything to the Magazines held in threadLocalMagazine.

Additionally, in Netty 5 i see no usedMemory counters: which will likely make this allocator to still perform better than the existing one in Netty 4.1.

Another pain point of the Netty 5 version is that allowing close means that users can really have fun creating many allocators, and per instance FastThreadLocal are a problem, because there's a giant array to back all the entries which will keep on being atomically increased as new instances are created....

[franz1981]

I think I forgot to close the magazine on ThreadLocal::onRemove

[franz1981]

Done @normanmaurer let's see what the CI think