Fix wetday frequency correction

[emileten]

• closes wet-day frequency correction uses only one value for each 'dry' data point #173
• tests added / passed
• docs reflect changes
• entry in CHANGELOG.md

essential changes :

• in the core function, instead of using a unique replacement value, use an array of (likely) different values, in process='pre'
• values strictly below 1 are modified, rather than below or equal to 1, in process='pre'.
• make tests deterministic (with a seed), smaller and more precise.

secondary changes :

2. in the core function, directly operate on the data array object within the dataset -- which means I have to introduce a var parameter -- default value is 'pr', but in tests we have other variable names. Need that parameter in the service as well.

[emileten]

@dgergel I requested your review for the method check.

@brews, in the computational side. I tested this on both Jupyter and Argo, and this ran within 3 mins using around 15GB of memory (that's because of the huge numpy sample array that I have to create, which has the shape of the GCM data). That fits in ~20GB resources for that step. We might need to increase slightly this number so that the ERA-5 data fits in.

But it will fit in the node and run fast.

[brews]

Thanks for cleaning this up, @emileten.

My one suggestion is to rename this new var argument to variable. This makes it consistent with the other functions and methods that grab a variable name. ...I see "var" and I think "variance" but that's an aside....

[emileten]

Thanks for cleaning this up, @emileten.

My one suggestion is to rename this new var argument to variable. This makes it consistent with the other functions and methods that grab a variable name. ...I see "var" and I think "variance" but that's an aside....

Thanks @brews, I changed this !

[dgergel]

This looks good, thanks @emileten for fixing this bug!

[delgadom]

the issue is here - changing from np.random.uniform(low, high) to np.random.uniform(low, high, size) changes the returned type from pure python float to ndarray[np.float64]. xarray coerces the former to ds[variable].dtype, but defers to the higher precision data type of the latter