Add an allow_nan keyword argument to dumps, defaults to True

[segfault]

with this ujson matches the builtin json behavior for NaN and Inf.
if a user wants to retain the old behavior they can pass allow_nan=False
to ensure strict json compatibility.

This should resolve #311

[hugovk]

Thanks for the PR! I think matching the stdlib json library is a good idea.

This is failing on Python 2.7, so let's merge it once 2.7 is dropped (#383).

I wonder, should this be considered a breaking change? It's okay if it is, as it'll be a major bump anyway when dropping 2.7.

And if so, do you think there should be a deprecation warning beforehand?

[tcopple]

Is there a timeline on dropping 2.7 support? Would love to have this merged in and released.

[hugovk]

There's been no objections to #383, and 2.7 is now EOL and the final 2.7 has been released, so I think we're good to drop 2.7 pretty much now.

#399 is the only thing in master as yet unreleased, I've left a comment there asking if it's needed in a final release supporting 2.7 first. If so, I'll do that first and we can then drop 2.7 shortly thereafter.

[hugovk]

#399 wasn't needed for Python 2.7, so I've merged #404 and we can make the next release Python 3-only.

[hugovk]

Thank you!

[Erotemic]

It seems weird that support for dumps was added, but no corresponding support for loads exists.

This means you can dumps things that you can't loads:

MWE:

    import ujson
    data = [float('nan')]
    text = ujson.dumps(data)
    print('data = {!r}'.format(data))
    print('text = {!r}'.format(text))
    try:
        recon = ujson.loads(text)
    except Exception as ex:
        print('ex = {!r}'.format(ex))
    else:
        print('recon = {!r}'.format(recon))