Improve matching API results

[pudo]

Some users of yente have reported issues with the matching API. These described problems are:

• Entities that are true matches score much too low, especially when the names are very short.
• Entities that have the same DOB are ranked very highly, irrespective of other features (like names).
• The way in which match is set to false when there are two matches is un-intuitive.
• Regulators want phonetic matching.

Here's some of the steps we're going to explore:

• Remove some sparse entities from the matcher training data
• Index soundex forms for all names
• Make name length less important in name match quality
• Implement a specific "OFAC style matching mode"

[pudo]

See: https://github.com/opensanctions/yente/releases/tag/v3.4.0