Do not store grids, destroy them after usage

[aulemahal]

Cleanest solution to #401 I could think of. An alternative less breaking solution would be to simply revert #387.

Once the weights are generated, we do not need the ESMpy grids (or locstreams) anymore. In fact, before #24, the grid were "destroyed" (they memory explicitly released by ESMpy) once the regrid was done. That meant that Regridder.grid_in was still there and with most properties accessible, but most methods would fail, with a segmentation fault. When implementing the Spatial Averager, I needed to reuse grid objects for multiple ESMpy regrid calls, so I commented out the explicit destroy.

As shown in JiaweiZhuang/xESMF#53, this creates a memory leak. #387 added these destroy back, as suggested by a comment on that thread. However, the moment they are executed seems to trigger #401.

Here, I make it clear that we do not need the grids : they are never set as attributes of the Regridder and are destroyed once the weights are generated. The destroy happens in the children (instead of a backend function) because that is were the objects were created. If users use the BaseRegridder with ESMpy objects directly, they should be unaffected.

But this is breaking because Regridder.grid_in and Regridder.grid_out do not exist anymore. Internally, we weren't really making use of them, but I don't know what are users are doing.