Improve machine_id ergonomics

[ZipFile]

I'm successfully using this library in production for a quite a while, but API around machine ids caused me some degree of discomfort. This PR fixes #8 as well as some improvements:

• Generating a lot of (millions) of ids efficiently is possible with multiple SonyFlake instances. Using callbacks to pass constant ids is counter-intuitive and not as performant. In this case you want to pre-generate set of random machine ids and iterate over list of id generators.
• Issue not generating unique id for single machine in celery tasks/threads. #8 mentions Celery, but the problem is much more general. If several id generators share the same IP, issue would re-appear. To some extent it can be mitigated by using random machine ids. On somehow related note, new UUID standard no longer recommends using non-random node ids which is essentially has the same role as machine_id in SonyFlake.
• Removed __new__ method, since it does not make sense. If some input params are invalid, None would be unpleasant surprise to encounter. Better be explicit and raise an error.
• Add option to pass now callback, for test reproducibility.

Overall, public API should be more-or-less backward compatible.

[hjpotter92]

Thanks for the PR. It looks really promising, and definitely something that can be published.