contour and contourf produce conflicing results for certain projections

Dieses Issue hat noch niemand übernommen.

Originally posted [here](http://stackoverflow.com/questions/33850052/matplotlib-why-inconsistent-results-between-contour-and-contourf) because I wasn't convinced it was a bug (I'm new to matplotlib)... but I'm increasingly convinced it is a bug, so here we are.

I'm plotting some NCEP temperature anomaly data using the stereographic projection. When I plot the data with filled contours, contourf, I see the first picture below. Note the red (positive values) over most of North America. When I change contourf to contour, making no other changes, I see the second picture below. Note the streak of blue (negative values) over central North America.
The contour result with the blue is what is actually in the data, and therefore what I expect to see. The contourf result is not what I expect. Is this a bug, or am I using these tools improperly? (Update: The first plot above should look like the below plot, which was produced using the cylindrical projection:)
The python code is reproduced below... you can obtain the data as a small binary file [here](https://onedrive.live.com/redir?resid=9F18EDDCF4E99E72!108&authkey=!ADXWQucUMyXgkJk&ithint=file%2cbin) if you want to try it yourself.

Some more notes... (1) shifting the longitude range to -180 to 180 using `addcyclic` and `shiftgrid` does not help. (2) the Lambert Conformal ('lcc') projection produces results that are _less_ wrong, but still not correct, when using contourf. (3) the Alberts Equal Area ('aea') projection centered on North America looks correct, however when I use the North Pole-centric version ('nplaea') I only see correct results for certain values of `lon_0`. I am happy to provide an example if this turns out to be a different issue.

```
import numpy as np
import matplotlib.pyplot as plt

# read data

f = open('data5.bin', 'r')
lat = np.fromfile(f,dtype=np.float32,count=73)
lon = np.fromfile(f,dtype=np.float32,count=144)
data = np.reshape(np.fromfile(f,dtype=np.float32,count=-1),(73,144))
f.close()

# plot

m = Basemap(width=10000000,height=6000000,
resolution='l',projection='stere',\
lat_ts=50,lat_0=50,lon_0=253)
m.drawcoastlines()
lon2d, lat2d = np.meshgrid(lon,lat)
x, y = m(lon2d,lat2d)
mymap = plt.contourf(x,y,data,levels=np.arange(17)-8,cmap=plt.cm.bwr)
plt.colorbar(mymap,orientation='vertical',shrink=0.75)
plt.show()
```

Beginne damit, das reproduzierte Python-Beispiel mit data5.bin auszuführen und contourf und contour unter der stereografischen Projektion zu vergleichen. Überprüfe, wie die projizierten x/y-Gitter an jeden Plot-Aufruf übergeben werden, und vergleiche anschließend die im Issue beschriebenen Ergebnisse für die zylindrische, Lambert-Conformal- und Albers-Projektion. Als erledigt gilt die Aufgabe, wenn contourf die erwarteten Daten einschließlich der Region Nordamerika abbildet, ohne die Eingabedaten zu ändern.

Vom Indexierungsmodell aus dem Issue-Text verfasst.