Extend compatibility to C implementation of frozendict

[H-Shay]

Fixes issue #36.
Background: some versions of frozendict contain a C implementation of frozendict that was incompatible with our code. This PR adds logic to catch and handle those cases.

[reivilibre]

I wonder if we should be checking frozendict.c_ext and choosing what to do based off of that (though I guess that attribute doesn't exist on old versions, so we would probably have to check it exists [getattr(frozendict, "c_ext", False)?])?
The use of exceptions for control flow is one of those things that is usually slightly nasty, though Python doesn't seem to shy away from it as much, so maybe I'm being too sensitive here.

[squahtx]

I wonder if we should be checking frozendict.c_ext and choosing what to do based off of that (though I guess that attribute doesn't exist on old versions, so we would probably have to check it exists [getattr(frozendict, "c_ext", False)?])?
The use of exceptions for control flow is one of those things that is usually slightly nasty, though Python doesn't seem to shy away from it as much, so maybe I'm being too sensitive here.

I did some incredibly unscientific testing and saw that the try-except was faster than doing a getattr when no exception is thrown (frozendict 1.2) and slower when an exception is thrown (frozendict 2.0.6 with c_ext). A try-except which doesn't raise seems to be just as fast as not using a try-except. We could determine which implementation to use at startup of course, which would be the best of both worlds, albeit more code to maintain.

IMO the PR as it stands is a definite improvement to the code and shouldn't regress performance for any of our existing supported frozendict versions.

[reivilibre]

I'm relatively happy to go with the above; it's hard to argue against the fact that it is a definite improvement (perhaps slow vs broken, with the existing case being pretty much unaffected).

We could determine which implementation to use at startup of course, which would be the best of both worlds, albeit more code to maintain.

For what it's worth, this seems like it would be a one-liner;

frozendict_using_c_extension = getattr(frozendict, "c_ext", False) and then using this in the if statement (if we're truly concerned about getattr performance — I'd be surprised if this was a particular concern because it should just be a hashtable lookup).

I feel a little bit bad about constructing a backtrace, throwing an exception and catching it for every frozendict we serialise, but it's not clear when the C extension is actually used (installing frozendict on my machine doesn't use it by default), and furthermore we're already resorting to copying the dict, which will already add its own performance overhead.

[squahtx]

frozendict_using_c_extension = getattr(frozendict, "c_ext", False) and then using this in the if statement (if we're truly concerned about getattr performance — I'd be surprised if this was a particular concern because it should just be a hashtable lookup).

Ah, good point. I was thinking about lifting the if out of the function entirely, which would be a few lines extra.