Make sure pearson_correlation_scalar input is validated

[JanisGailis]

Input validation in pearson_correlation_scalar had faulty logic

[JanisGailis]

Closes #746

The pearson_correlation_scalar is meant to work only on 1D variables (much like the underlying scipy function). I had a check for 3D variables, which was faulty. This PR resolves that, but the operation you were trying to carry out in #746 will then fail anyway, just with a nicer message.

[codecov-io]

Codecov Report

Merging #753 into master will not change coverage.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master     #753   +/-   ##
=======================================
  Coverage   76.94%   76.94%           
=======================================
  Files          81       81           
  Lines       12498    12498           
=======================================
  Hits         9616     9616           
  Misses       2882     2882

Impacted Files	Coverage Δ
cate/ops/correlation.py	100% <100%> (+1.06%)	⬆️
cate/util/process.py	90.15% <0%> (-0.76%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9af8fd2...a044555. Read the comment docs.

[JanisGailis]

@forman Merge when ready. If this is not merged in a day, I'll go ahead and do it myself.

[forman]

Why is that operation 1D only? And why specifically about time series?

Can't we just check if both datasets have same dims and all dims are equal. If so, flatten both nd-arrays and compute correlation.

[forman]

@JanisGailis merging this, but please answer my question once you have time!

[JanisGailis]

@forman Using .values.flatten() on an ND xarray.DataArray, is a bad idea, as it will very likely run into MemoryError in many instances.

However, after some quick investigations, xr.DataArray.stack() could be used to achieve what you suggest in a memory safe way. This probably means that the datasets will have to be coregistered beforehand, for the operation to make any sense on ND inputs. This shouldn't take too long to implement. Should I go ahead?

[forman]

I actually did not mean the numpy flatten(), but just flatten, turn ND into 1D. So, yes, fine!