Support for multi-gpu or cpu-only training

[natlouis]

Reference to #50

Adding new config parameter: num_gpus (default: -1)

-1: Use all available GPUs
0 : Do not push model onto GPU i.e. use CPU
>1: Use specified number of GPUs

• Automatically loads weights regardless of whether it came from single or multi-gpu training.
• Config batch size is multiplied by the number of GPUs