Tweaks for monthly mean calculation.

[simonpf]

Modifications to the monthly mean calculation to speed up the processing on sun.

Main changes:

• Use glob instead of rglob: Recursive glob can take a long time in a directory containing many .zarr files because each zarr file is itself a directory.
• Sparse accumulation: Reindexing files to the all time steps of the day is slow and requires a lot of memory. Replaced it wich accumulation that uses only the corresponding time steps.
• Data types: Use single precision for accumulation, which I think should be sufficient. However, probably better to make this a flag to allow checking its impact.
• Add option to parallelize processing of multiple months.

The changes took the processing time for a single month down from about 20 to 10 minutes on sun. Moreover, the reduced memory requirements allow running about 8 processes in parallel.

[simonpf]

I compared the means calculated using single and double precision (see plot below) and it does not make a significant difference. So it seems safe to run the calculations using single precision.

[adriaat]

I compared the means calculated using single and double precision (see plot below) and it does not make a significant difference. So it seems safe to run the calculations using single precision.

What does the plot show? In any case, I also think single precision should be OK, but recently (working with another problem) using single and double precision when computing means from a large array did make a difference. So if we can allow it, I'd compute the means with double precision.