Changement de la parallélisation des traitements avec rioxarray

[cadauxe]

Fait référence à l'issue #12 .

[cadauxe]

Une option lors du lancement des tests dans la CI a été rajoutée, on peut maintenant accéder au temps que prend chaque test et comparer si tests mettent plus de temps avec rioxarray qu’avec la classe Windowable.

Voici les conclusions majeures :
test_timeseries_command_line_default : de 25.13s à 58.60s (+33,47s)
test_filtering_command_line_default : de 14.15s à 10.24s (-3,91s)
test_radioindice_command_line_default : de 4.71s à 9.05s (+4,34s)
test_zonalstats_command_line_categorical : de 6.49s à 8.50s (+2,01s)

[sylvesterkaczmarek]

This PR includes generated/runtime artifacts such as __pycache__/*.pyc, .coverage, *.egg-info, METADATA/RECORD/WHEEL, docs/_build, and generated test outputs. These are environment-specific and make the 1,992-file diff non-reproducible. Please remove generated artifacts and keep the source/config/test changes only.

[sylvesterkaczmarek]

Hi, the new valid/nodata accounting is off: all_count = np.count_nonzero(~mask) already counts only unmasked pixels, which is the same population as dataset.count(). That makes nodata always zero and valid essentially 1.0 even when the clipped region contains nodata. Could all_count represent all pixels in the geometry instead?

[sylvesterkaczmarek]

One band-selection regression: both compute_speed() and _interpolate_xarray() use isel(band=slice(0, len(bands))), so the values in bands are ignored. For example, bands=[2, 4] processes the first two bands rather than bands 2 and 4. Could these paths select the requested band coordinates/indices explicitly?

[sylvesterkaczmarek]

Hi, multi-band category stats are collapsed here. roi_statistics is one dict shared across every band, so entries for a given category/stat are overwritten by the next band, and the function always appends a one-element list. Could the per-band dictionaries be kept separate as in the previous return contract?

[sylvesterkaczmarek]

One empty-region regression: the old code returned None statistics when dataset.count() == 0, but that guard is gone. With majority or minority requested, an all-nodata clipped region produces an empty pixel_count and _key_assoc_val() calls max([])/min([]), raising instead of returning an empty statistic. Could the zero-valid-pixel case be restored?