Improved fooling of IPython's hard-coded backend-to-gui mapping.

[anntzer]

@afvincent I think this should fix the other half of #3 for you (basically ipython was essentially forcefully re-fetching the backend from your matplotlibrc (which uses qt5agg) rather than respecting MPLBACKEND (which should have priority)); can you confirm?
(IPython uses a hard-coded list of interactive backends to gui toolkits mappings and obviously mplcairo is not in there. So I just need to reverse-engineer the best way to fool it... In fact I just saw that I had noticed this specific issue a long time ago, ipython/ipython#10255.)
It's pure python so I'm not going to bother making a new wheel -- just patch mplcairo/base.py yourself :-)

[afvincent]

After replacing mplcairo/base.py with this version from this PR, the issue seems to be fixed indeed :):

In [1]: run example_issue_mplcairo.py
#0 module://mplcairo.qt
#1 module://mplcairo.qt
#2 module://mplcairo.qt
#3 module://mplcairo.qt

Thanks @anntzer !