Incremental Torch improvements

[SergeyTsimfer]

• refactor BasePool: it can be split into multiple classes with clear functionality (done in pooling refactoring #469)

• improve Encoder/Decoder modules

• refactor n_iters and decay configurations: no need to pass n_iters in the root configuration (done)

• make so every block is sent to device: can be helpful with pre-trained models (done in Torch improvements #461)

• refactor pyramid layers so they use common base

• make Xception out of XceptionBlocks

[nikita-klsh]

• not specifying layout in the initial_block / body / head of TorchModel config results in skipping module at all (done in Torch improvements #461)

[SergeyTsimfer]

• add some warnings if task is classification and classes are not set in config: in this case, if model is built off of the first batch with data, classes can be infered incorrectly (understandably so)

[SergeyTsimfer]

• make better name for _add_output_op in case of callable operation (done in Torch improvements #461)

• allow kwargs for TorchModel.save method (done in Torch improvements #461)

[nikita-klsh]

• kwargs for combine passed through config in Decoder do not reach the Combine module inside (done in Torch improvements #461)

[SergeyTsimfer]

At the time of implementation, we've decided to keep the argument names of PyTorch layers the same as their TensorFlow counterparts. For example, dilation_rate instead of dilation; strides instead of stride. I think it is time to fix that and either change names to the PyTorch convention or make methods recognize those parameters as aliases

[SergeyTsimfer]

Also, looks like we should change the letter of upsampling to U with parameter mode, and, maybe, P for pooling operations; otherwise, it is confusing to use nn interpolation which is also n letter (same as normalization layer)