Add two popular datasets for character level LM

[entron]

Added data preparation and example trainning config scripts for two popular datasets: text8 and enwik8.

[entron]

Also tried out feeding the outputs of each layer to itself mutiple times in the 2nd commit.
For the shakespeare_char dataset, this actually gives better val at 1.4543 with only 1 layer and 1.8M parameters.
For bigger datasets such as text8, this also gives better results when the
number of parameters are the same. Haven't tested on GPT-2 yet.
The 2nd commit may be not so relavant though.

[karpathy]

That's nice, but prefer we keep n_layer_update separate

[entron]

I have removed the 2nd commit.

[Andrei-Aksionov]

Maybe it's about time to have a separate .py file with the shared logic?
Because all prepare.py files for shakespeare and these two new datasets basically do the same thing.
I understand that it's sometimes better to have some code duplication for the sake of simplicity and easiness of understanding, but this is not the case (in my opinion).

I am opened to hear why I am wrong (again 😄 ).