Faster rl.{first|last}_run implementation

[coxipi]

Generic Issue

Sub-function _boundary_run uses rle, but it could use _cumsum_reset_on_zero, effectively avoiding these two steps in rle:

cs_s = cs_s.where(da.shift({dim: -1}, fill_value=0) == 0)
out = cs_s.where(da == 1, 0)  # Reinsert 0's at their original place

These steps are probably much less demanding than the cumsum operations, but still, it's worth considering. I think there could be merit in keeping things as they are if we feel it makes the code more understandable, but I'm not sure about this, and also the run_length functions are somewhat at a low level in the code, so I have the feeling that performance should be prioritized here.

Side note about return type of first_run

I currently have to convert the result of first_run to be able to use it for time selection. There is no reason really for the output not to be already a int (of course, in the case where coord is False or None)

end_season_index = rl.first_run(sim==0,window=14).astype(int)
mask = xr.where(sim>0, 0,np.NaN)
mask[{"time":end_season_index}] = 1

This could be fixed by adding a else statement in coord_transform:

if coord:
    crd = da[dim]
    if isinstance(coord, str):
        crd = getattr(crd.dt, coord)

    out = lazy_indexing(crd, out)
else:
    out = out.astype(int)

Code of Conduct

• I agree to follow this project's Code of Conduct