Hunspell stemming problem [LUCENE-6175]

I found j-morph which is an java implementation of hunspell. Here it is: https://code.google.com/p/j-morph/ I know it has a terrible source code but somehow it works much better than Lucene's hunspell implementation.

I used the same hungarian aff and dic file (j-morph contains it) but j-morph produced much better results.

I will be really impressed If it could work as good as j-morph does. Now sadly I cant accept current Lucene's hunspell implementation. I really like hunspell please improve it as much as possible :)

---
Migrated from [LUCENE-6175](https://issues.apache.org/jira/browse/LUCENE-6175) by Baló András
Environment:
```
All
```

Start by reproducing the reported stemming difference with the Hungarian .aff and .dic files used by j-morph, then compare Lucene's Hunspell results with j-morph's output. The issue names no Lucene files, tests, or entry points, so locating the relevant implementation and defining acceptable parity would require project familiarity. Done means Lucene produces materially comparable Hungarian stemming results.

Written by the indexing model from the issue text.