Actual ngram features depend on field entry pointers, so we can't change them. We can completely rebuild trie index and dictionary entry table.

Logic should be:

1. Do the step 1 of dictionary import (compute unique column values with frequencies)
2. Read the present string storages and remove already stored values from the step 1 results.
3. Append new values to each string storage
4. Rebuild entry table.
5. Rebuild index.

Probably, we should implement the initial import as append to an initially empty dictionary to have only one implementation of csv dictionary import logic.

The open question is what should we do with full duplicates.
Should the import logic perform the full deduplication of the dictionary entries in addition to the current logic?