Improve pinyin fuzzy segement algorithm

[wengxt]

Previously, we blindly choose the segment to always prefer the longer
next match, this is prove wrong in the case of "sangeren".

Which should produce, "san ge ren", "sang er en", "sang e ren".

Instead, we change the check to be:
if (current + next match) is valid, and complete pinyin, make it an
acceptable option, unless (current, next match) is actually an inner
fuzzy, which is handled separately below.

For example:

1. For sangeren, will produce sang & san, since next match of
   "san", which is "ge", is a complete pinyin.
2. For hua, will only produce hua, since hu a is a inner fuzzy.

Even if it will produce "extra" segement, for example, in the case of
"sanger" will produce a partial pinyin "san" "ge" "r". We may still
consider it as make sense. Since partial pinyin match is considered
fuzzy and will have a penalty score.

People may even benefit from such segement, since "san ge r" seems to be
the most possible option.

Fix #87