Check how NaN is propagated through CSET

[jwarner8]

We very soon have mask capability in CSET which will replace data with np.nan where data is masked out. For subsequent operators (as masking is unlikely to be the only operator used in a recipe), we should check that other operators can handle these (e.g. domain averages do the appropriate nanmean).

[daflack]

NaN propagation was an issue in #1874, but a fix has been implemented in the collapse operators - not sure if it needs to be applied to other operators - may be worth reviewing after #1874 is closed.