ONNX runtime support

[joeyballentine]

• Implements ONNX runtime inference
• Allows converted PyTorch models to be chained into the ONNX load model
• Re-implements pipInstallWithProgress as this feature was removed and is useful (can be replaced in the future if a different solution is found)
• Implements the ability for dependencies to have multiple packages associated with them. For ONNX, this includes onnx, onnxruntime-gpu, and protobuf
• Fixes bug with auto-split-upscale max_depth passing, which would cause split depth caching to break if it was actually implemented (I fixed this in my esrgan fork a while ago)