python garbage collector fail when tile processing with multiple device

[StRigaud]

I just pushed a demo notebook for tile-processing accross multiple device using dask and zarr.

It actually work well but with a major issue, it does not seems that the garbage collector is working in the dask system. Hence, each tile is being push in the device memory but not released when finished processing. In the example the data fit in my GPUs but this example would fail on a very large data, which kind of fail the reason why one would use such approach.