backend_pgf: clip paths within the backend (fixes #1857)

[pwuertz]

Because LaTeX doesn't handle very large numbers (#1857), this PR moves some of the clipping operations from the pgf-code into backend_pgf.

[pwuertz]

Weird, python 2.6 and 3.x tests are passing on Travis, 2.7 is not (although it passes on my machine with 2.7). It also won't show me the log-file to see what's wrong :/

[mdboom]

It looks like the Travis failure was a networking error when trying to install PIL, so in other words a false result. Looks like a reasonable fix to me.

Should this perhaps go on 1.2.x since it's a bugfix? Also, it could use a test -- I don't know if it needs to be an image comparison test, but something to just make sure the really large number doesn't show up in the output should be sufficient.

[pwuertz]

I just added a test that triggers #1857 and now passes with clipping added to draw_path.

Also, apparently some restructuring has been applied to master and a failed LaTeX call now results in an error looking up subprocess.CalledProcessError. I forwarded the missing Exception in matplotlib.compat.subprocess.

If this is fine with you we can apply this to master and I'll prepare another PR for 1.2 without the subprocess fix.