Model builder changes

[ASvyatkovskiy]

The PR implements a set of changes to standardize runner and mpi_runner modules, affects the model builder module.

1. Compile models in runner and mpi_runner. Do not recompile the model in mpi_runner
2. Only build NN architecture (layers, activations) in the model builder, add optimizer/loss later (in runners)
3. Provide more optimizer options, including TFOptimizer

Minor changes:

1. Do not mask all GPUs via CUDA_VISIBLE_DEVICES for architectures with only 1 GPU per node (OLCF Titan)
2. Minor changes in the conf: enable learning rate decay, but reduce patience.
3. Reduce base learning rate as a function of number of workers to restore model convergence