Changed Pearson r to handle masked data in a more reasonable way

[niallrobinson]

Peaerson r now only considers a subset of the data where corresponding data in both cubes are not masked. This has implications for the assumed length of the cubes, and can change the calculated r values

in response to #1534

[marqh]

just a typo *masked

[ajdawson]

I found the logic a bit hard to follow on first read. I'd personally prefer something a bit more concise, I think the following is equivalent to lines 101-120:

a_mask = ma.is_masked(reshaped_a) and reshaped_a.mask
b_mask = ma.is_masked(reshaped_b) and reshaped_b.mask
mask_union = a_mask | b_mask
if not np.isscalar(mask_union):
    return_a = reshaped_a.compress(~mask_union)
    return_b = reshaped_b.compress(~mask_union)
else:
    return_a = reshaped_a
    return_b = reshaped_b
return return_a, return_b, res_ind

I guess there could be arguments against this too, but at least there are fewer lines of code to understand.

[ajdawson]

@niallrobinson - This work would be superseded by the changes in #1714. The changes there make handling masks and weights a more manageable task.

[ajdawson]

I'm closing this since the whole pearsonr code has subsequently been re-written.