Add advanced validations for dictionary building

[eiennohito]

We plan to introduce dictionary build warnings, which will not abort the building of the dictionary, but will report that something was not good.

Warning-producing checks will be optional, but enabled by default.

Proposed list of warnings:

• Surface forms are not normalized. Words with such surfaces will not be possible to lookup via Trie index (this is current behavior) and those problems seem to appear somewhat frequently with user dictionaries.
• Word segmentation producing non-consistent splitting. Concatenation of word splitting surfaces should produce the surface of the original word.
• Having non-distinguishable dictionary entries (with same left/right connection IDs + surface). In this case an entry with the highest cost wins, otherwise the last dictionary entry wins. We will remove all other entries from index.