Refactoring of the SPNG::DNNROI

[physnerds]

Currently the DNNROI does data-preprocessing, inference and post processing. Inference is done via TorchService.
Refactor DNNROI such that DNNROI is agnostic to the data model. Pre-processing steps like downsampling and post-processing steps like upsampling should be done by a separate interface.
Look at issue #22 for the design discussion of these interfaces.
Look at issue #19 for the Data model standardization.

[brettviren]

For this, a new node, say TorchForwardNode would merely marshal one ITorchTensor from input, through the ITorchForward and result to output. Some minimal pre/post processing would be limited to how to select the input ITorchTensor (perhaps by configuring a datapath regex string) and how to place the result (perhaps by configuring a datapath format string). Then, all pre/post processing that is in the current DNNROI node would be moved in to one or more generic function nodes.

[physnerds]

DNNROI is refactored into the PreProcessing, Inference and PostProcessing as shown in the link here:

• Preprocess
• Postprocess
• Inference
• Validation of Implementation