read EAM native spectral-element grids, and make formats pluggable

[berkgeveci]

QuickView handled one grid: the ne*pg2 physics grid, whose SCRIP connectivity gives every cell its own four corners and whose values are cell averages. The dynamical core runs on GLL nodes shared between spectral elements, and its values are nodal point samples -- the opposite shape. Supporting both means the format, not the application, has to decide what the pipeline looks like.

pipeline.py now has a DataPath base with one subclass per format, chosen by sniffing the connectivity file (element_corners vs corner_lat):

Pg2Path cell data reader > CenterMeridian > Extract > Project > Surface
DycorePath point data reader > CenterMeridian > Extract > Project > Surface

EAMDycoreSource is the tier-1 reader: one bilinear quad per (np-1)^2 subcell, vertices are the GLL nodes, so no topology is reconstructed and no point is invented. It lays the sphere flat itself, duplicating nodes at the date line and the poles rather than clipping, which keeps every cell whole and every value exact. HOMME winds element_corners inward, so the winding is reversed at read time; a pole node's longitude is arbitrary and must take no part in the seam test, or float noise flings a corner a whole turn out of the map.

The map's longitude window is now movable: the left edge sits anywhere in [-180, 180] and the right edge a turn east of it. One rule shared by every filter -- cut at c = input_origin + (origin - input_origin) % 360, translate the halves by origin+360-c and origin-c -- and at origin -180 with a [0, 360) input it reduces to the previous hardcoded behaviour. EAMCenterMeridian needs to be told its input's window because the two readers emit different ones. The projection is re-centred before PROJ sees it: PROJ normalises longitude into [-180, 180) before applying lon_0, which would send the window's right edge to the left rim and draw cells and coastlines across the map.

Association is plumbed through view_panel, compute.extract_avgs and app.py, and ViewManager.drop_views() discards views on a format switch, since each binds its mapper to the tail of one specific pipeline.

Performance: EAMCenterMeridian passes through untouched when the input already sits in the requested window (the dycore default), and on a real rebuild hands the cached points and cells objects downstream, because EAMProject keys its cache on point identity and EAMExtract on their modified time. EAMExtract keeps its cache for nodal data through a point pedigree map -- it removes whole cells and never interpolates, so the ids stay exact. Roughly 7 ms per pipeline pass at ne30 with 20 variables loaded, matching pg2.

Verified against ParaView 6.0.1 (VTK 9.5.2) and ParaView master (VTK 9.7).

[berkgeveci]

@jourdain can you review?

[jourdain]

I didn't see where that is getting called, but having it seems good for a follow up integration.