Improve performance in multi-threaded synchronized case.

MultiArena allocator is now faster than the default system allocator also in synchronized multi-thread mode.

Improve cache locality in synchronized mode

Synchronized mode is now a bit faster because allocation and deallocation counters are always living in the same cache line.

Fix zero-sized allocations and `nullptr` deallocations

Fix zero-sized allocations and nullptr deallocations.
Also some compiler warnings removed.

Fix upstream allocator propagation.

Merge pull request #2 from tirimatangi/fix-allocator-propagation

fix allocation propagation

Initial release

Initial release.