Many changes to make the program more efficient

This version is about 2 to 3 times faster.
* State in the HMM is not a Named Tuple anymore but a normal tuple
* Tags (POS Tags and Morp-Tags) are not striings anymore but integers. Two dictionaries are added to the model to translate between strings (in input and output) and internal integers
* For any tag X the corresponding tag END_X is obtaine by multiplying with -1
* If Viterbi is run with a target POS, we use a precompiled list of states that could lead to the corresponding tag
* Many small improvements