Move pytorch-based models over to TorchServe

[lucaro]

Whenever we use external models for feature extraction that are based on torch, it's presumably much more efficient to serve them through TorchServe.

[ppanopticon]

So the tasks here is to create an abstract implementation of an Analyser (and its Retriever and Extractor) that can be used with TorchServer?

[lucaro]

Pretty much, yes. TorchServe also has a gRPC API that might be more efficient than having to do things via REST. I have not yet looked into the details of it though.