Profiling and accelerating learning

[bamasa]

Пробежался профайлером (результаты в profiling_results.ipynb). Вот какие идеи появились (+некоторые мелкие вещи исправил):

• "cdf" will go away with replacement "normal" approach by "Hutor" one. So we reduce time approximately by half.
• we could gain some time by optional exceptions using in api, like this:

  if using_excetions == True:
      test_sth

  end disable testing during ML.