Prevent memoy allocations in Filterer

[DavidDureau]

Today, we can use only once an instance of Filterer.

The Filterer implementation allocates temporary arrays and it can be expensive especially with the CUDA accelerator runtime.

In order to prevent theses temporary allocations, is it possible:

• to reuse a Filterer instance with the same temporary arrays;
• and/or to use memory pools (Add LLNL/Umpire backend for IMemoryRessourceMng #112, Using memory pool to reuse allocated arrays and variables  #933).