improve performance of interpolate edges/nodes to faces

[veenstrajelmer]

2D indexing with face_node_connectivity or face_edge_connecitivity is terribly slow: dask/dask#10237
This is used in dfmt.uda_edges_to_faces() and dfmt.uda_nodes_to_faces(), also check whether we use it more often.

Faster example with stack and unstack:

import xugrid as xu
import xarray as xr

ds = xr.open_dataset("conv_schism_raw_2.nc").chunk({"time": 1})
grid = xu.Ugrid2d.from_dataset(xr.open_dataset("ugrid.nc"))
node_salinity = ds["salinity"]

reduce_dim = "mesh2d_nMax_face_nodes"
indexer = xr.DataArray(grid.face_node_connectivity, dims=(grid.face_dimension, reduce_dim))
valid = indexer != -2
indexer = indexer.stack(__tmp_dim__=("mesh2d_nFaces", "mesh2d_nMax_face_nodes"))
face_salinity = node_salinity.isel({"nSCHISM_hgrid_node": indexer}).unstack("__tmp_dim__")
result = face_salinity.where(valid).mean(reduce_dim)

Maybe this is a good moment to implement in xugrid instead of dfm_tools.

Other open vindex issue: dask/dask#7313