Optimize wrappers

While our wrappers avoid the extra process overhead, I think they can still be improved.

e.g. we only need to run https://github.com/apertium/lttoolbox/blob/master/python/lttoolbox.i#L23-L24 one time for any given `Analyzer` object

On another note, we should also try to do something smarter instead of re-opening files everywhere (but this might be harder to get working).

@unhammer, is doing the first thing I mentioned (i.e. only loading the dictionary once) meaningful in terms of improving performance? Sounds to me like it would be.