Make interpolate_to_grid more robust

[dopplershift]

There are some gremlins lurking in interpolate_to_grid, at least with this dataset:

import metpy.interpolate as minterp
from metpy.io import add_station_lat_lon
from siphon.simplewebservice.iastate import IAStateUpperAir

df = IAStateUpperAir.request_all_data(datetime(1993, 3, 13, 12), pressure=500)
df = add_station_lat_lon(df)
df = df.astype({'latitude': float, 'longitude': float})
df = df.dropna(subset=['latitude', 'longitude'])

methods = ['rbf', 'barnes', 'cressman', 'linear', 'natural_neighbor']
fig, axes = plt.subplots(nrows=2, ncols=len(methods), figsize=(15, 8))
for ind, ax in enumerate(axes.flat):
    method = methods[ind % len(methods)]
    project = ind >= len(methods)
    if project:
        proj = ccrs.LambertConformal()
        x, y, _ = proj.transform_points(ccrs.PlateCarree(), df.longitude.values, df.latitude.values).T
        res = 75000.
    else:
        x = df.longitude.values
        y = df.latitude.values
        res = 1.0
    gridx, gridy, grid_height = minterp.interpolate_to_grid(x, y, df.height, interp_type=method, hres=res)
    ax.imshow(grid_height)
    ax.set_title(f'{method}, project={project}')

In my opinions, we need to try to make everything operate sensibly default out of the box. Cressman seems ok here, but I swear I saw weirdness elsewhere. Barnes and RBF are not good here.

1. Investigate what's going wrong with RBF. Also, internally we are using the Rbf class rather than the newer RBFInterpolator
2. We already knew the way Barnes was formulated here was problematic, but we really need to fix this up.
3. Definitely want to make natural neighbor more reasonable performance-wise (Port natural neighbor interpolation to Shapely #2622)

[ThomasMGeo]

Another implementation of Barnes Interpolation in Python with a nice paper.