Towards TF dataset, part III

[jindrahelcl]

• move to TF 1.12
• switch to Python 3.6
• remove check_dataset_and_coders (this did not work properly for some time, will be obsolete with tf dataset)
• refactored execution result objects
• refactored experiment and learning utils
• placeholders constructed outside model parts (a step towards having a separate tf dataset tensors)
• batching is now a property of the dataset, not training. batch_size remains in [main] for backward compatibility
• added some tests
• image reader and numpy readers do not infer some properties to be able to provide explicit shape information to tf dataset