Use Levenshtein distance from Jellyfish library

[tomjaspers]

I've created a small benchmark where I generated 10 misspellings (using a probabilistic approach) for all the artists / albums / songs in my library (about ~5600 items).

These are the results comparing 3 popular C packages, with our current one

python-Levenshtein
  elapsed time: 0.0654819011688
jellyfish
  elapsed time: 0.0748250484467
editdistance
  elapsed time: 0.19926404953
current one: handrolled (from Wikibooks) 
  elapsed time: 12.266146183

Despite being pretty informal benchmarks, they do show the two orders of magnitude difference.

I've decided to go with jellyfish over python-Levenshtein because:

• python-Levenshtein is no longer actively mainted
• python-Levenshtein crashes if two strings are different types (e.g., str and unicode)
• the performance difference between the two is neglectable, esp. considering our current one
• Jellyfish works on Windows (it falls back on a Python implementation if the C extensions can't be compiled, which is the same as our current one)

---

In #646 it was brought up to use the package optionally (and fallback on our own implementation if it wasn't installed). I don’t really see a reason for this, considering Jellyfish works everywhere and does this already.

Oh and the refactoring plans mention something about using difflib.ratio(), but it is in fact not a Levenshtein distance.

[sampsyo]

Fantastic! I'm all for it. And I didn't know that Jellyfish had a fallback to pure-Python built in—since that's the case, you're right we don't need our own fallback.

Since the next release is already huge and just around the corner, let's hold off for a couple of days and make this the first change in the next version.

[sampsyo]

OK, version bumped! Merging this now. Thanks again!