Fundamental change in sel and seltime methods - now not masking NaN-v…

[knutfrode]

…alues, as these from now on will have significant meaning as dividing trajectories into segments. to_1d() is now also not removing NaN-values.

[knutfrode]

@gauteh It would be good to have your check of this before proceeding.

example_concat works locally, so not sure why it fails here.
The failing line 54 is dc = xr.concat((d1, d2), dim='trajectory', join='outer'),
but neither d1 nor d2 contain NaN (holes) or duplicates, although they are created with to_1d that does now allow NaN

[knutfrode]

Inserting these two lines before 54 (concat) solves the Doc-build problem. But this is quite strange, since no duplicates are found locally.

d1 = d1.drop_duplicates('time')
d2 = d2.drop_duplicates('time')

[gauteh]

What if obs_dim is unordered? I think this has to work on indices, not the values of the obs_dim. I guess obs_dim is time since this is 1d, but comparing floats is not so safe. To find the first/last value matching argmax is better.

[gauteh]

Feels like we have to try this one out for a bit, I guess we could merge and potentially revert... It's nice to be able to use standard sel etc in 1d, so it's a bit of a pity if that cannot be done anymore. Maybe sel should check if there's any NaN's before doing the special version.

Now you're slicing on first and last match, but sel doesn't need to always be a slice. It can be an arbitrary list. So the new sel is more of a slice than a sel :)

[knutfrode]

You are of course right, I simply forgot other options than selecting a slice...
Will have another look.

[knutfrode]

@gauteh Now traj.sel is simply using self.ds.sel as before (NaN not accepted), whereas logic to preserve NaN is moved to traj.seltime

One note: for several methods that e.g. padded/added items along the obs-dimension I have now re-written the obs-dim coordinate as integers from 0 to N before returning.
Earlier there were duplicates (e.g. [1, 2, 3, 4, 4]) or NaN-values.
However, this then also has to be consistent, otherwise it is not possible to compare arrays with differing indices. Thus I had to insert this also at methods where it was (apparently) not necessary before, e.g. here: f212ee9