wordfreq 1.4: some bigger wordlists, better use of language detection

[rspeer]

Here are the major changes I've made in this branch:

• I added "large" lists, containing word frequencies down to 10^-8 instead of 10^-6, for languages with enough data to justify them: English, French, German, Spanish, and Portuguese.
• I revised the language detector that applies to Twitter and Reddit. Before this, Reddit language detection was being handled particularly poorly, with for example spam in Chinese being treated as English.
• I removed Reddit comments with scores below 1, decreasing the effect of spam and interminable flame-wars.
• After routing language-detected Reddit comments to the correct languages, I kept the content in English, Spanish, German, and Swedish, four languages that seem to be well represented on Reddit. (Some other languages, such as Arabic, French, Italian, Portuguese, and Russian, have too much spam.)
• I added the ability to get frequencies on the "Zipf scale", which Brysbaert wanted me to add in exchange for using his data.

[alin-luminoso]

So we do expect a science output change from this, right? Do we need to benchmark it?

[rspeer]

You said that there were some small changes in Russian. Can you give me an example of something that changed? The top 1000 words appear to be identical to 1.3 in order, and I checked a couple and found that their frequencies are equal as well.

[rspeer]

Okay, the changes come from disregarding Tweets that are shorter than 50 characters and detected as non-English.