Add ds.traj.filter() for outlier removal (speed and n-sigma sliding)

[gauteh]

Implements filter(method, ...) on the traj accessor:

• 'speed': masks positions where speed to next point exceeds max_speed [m/s].
  Uses per-step time differences (not a fixed scalar) so it works correctly
  on non-uniformly sampled drifter data.
• 'nsigma_sliding': masks lat/lon where either deviates more than nsigma
  standard deviations from a sliding-window local mean (same algorithm as
  sliding_filter_nsigma in readers/omb.py, verified by test).

Implementation follows the Traj1d/Traj2d delegation pattern:

• _nsigma_sliding_filter() helper in traj1d.py
• Traj1d.filter() contains the logic
• Traj2d.filter() delegates via trajectories().map(d.traj.to_1d().traj.filter(...).traj.to_2d())
• @AbstractMethod stub with full docstring in Traj

Also:

• Add Dataset.traj.filter to api.rst
• Add filter demonstration section to examples/example_drifters.py
• Fix examples to unpack (paths, ax) from traj.plot() after plot.lines()
  was updated to return both values
• Add trajectories().map() delegation pattern to copilot-instructions

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[knutfrode]

This filter method looks useful and convenient.

However, this is a fairly specific method (though with general name) that partly overlaps with more generic and flexible functionality such as ds.traj.insert_nan_where(<custom condition>)
https://github.com/OpenDrift/trajan/blob/main/examples/example_drifters.py#L60
But as far as I can see, it looks like this filter method is different and more sophisticated compared to insert_nan_where, and therefore justified ?

In principle this filter method could be separated in two:

1. specific method to provide indices of positions to be filtered
2. generic method to apply filter based on conditions from 1) or other custom conditions

[gauteh]

Yes, I first tried to use ds.traj.speed() to remove jumps, but it didn't work very well. Then there's the nsigma filter from @jerabaul29's reader, which was only available to the OMB reader.

It seems that there are specific edge cases which are difficult to combine the basic operations into, and then a bit difficult to remember how to use. So I think it would be a useful method to many, and more intuitive to remember than combining the other methods? Especially, since it is now more advanced than what is easily possible by combination of the others.

I agree that it could be split, perhaps it could also (in the future) take a custom method or list of indices as it's filter, but I think it would be useful to keep both speed and nsigma as built-in. It may be that the nsigma one should be the default one. Right now I'm actually using both: ds.traj.filter().traj.filter('nsigma_sliding') and it works very well.