Allow base length parameter for ngrams

[rgertenbach]

Right now models are trained on which n-1 words follow the last word.
You should be able to train which n-base words follow the last base words.

word1 word2 word3 word4 word5
  1     2     3     4     5
  B     T     T     T     T  // Base 1
  B     B     T     T     T  // Base 2
  B     B     B     T     T  // Base 3
  B     B     B     B     T  // Base 4

Requirement: 1 >= Base < n