Reduce memory footprint of the conversion

[emmanuelmathot]

During the conversion process the dask process memory increase quickly due to the full copy of the source data. Once copied, largest data group such as r10m and r20m are not needed anymore. It would be valuable to clear the chunks from the memory to reduce the memory footprint.

[d-v-b]

for downsampling, is it acceptable to use the (N-1)th scale level as the input for the Nth scale level? In this case, we can keep memory usage very low because we can read the result of the previous downsampling from local storage.

[emmanuelmathot]

120m is made of r60m group, and so on for 360 and 720 so keeping r60m is sufficient