Port to tensorflow 2.3.0 and refactor

[elazarg]

Main changes:

• Refactor and simplify architecture code
  • Refactor common architecture code to dedicated functions
  • Refactor experiment code to a single function
  • Add titles to to the plots
• Don't use keras directly; use tensorflow instead (keras is not independent of tensorflow anymore).
  • No need to cleanup sessions, so keras.backend is not needed.
• Create directory when it does not exist (Fix enable smooth reproduction #4)

The functions for word and for char are not always exactly the same. I wonder if that's intentional.

The code for BiLSTM crashes on my machine, so there are no plots there, but this has nothing to do with this PR.

(This PR probably works with 2.2 too)