[Feature] TorchServe Integration

[ppanopticon]

This PR provides features based on TorchServe. It consists of the following components:

• An abstract TorchServe analyser that provides the basic functionality to interface with TorchServe via gRPC. This can be used to create concrete implementations for specific models.
• A concrete TSImageLabel implementation that facilitates labelling of images. This can be used with most of the classical inference models that can generate labels from images. The implementation can serve as an example for other variations.

A TSImageLabel feature can be configured as follows:

"fields": {
 "torchserve": {
  "factory": "TSImageLabel",
   "parameters": {
    "host": "127.0.0.1",
    "port": "7070",
    "model": "densenet161",
    "token": "099-Ed4q",
    "threshold": "0.5"
  }
 }
}

Important: It is up-to the user to deploy the model in TorchServe beforehand. The following parameters exist for configuration (the first four are used irrespective of the model that is being used):

Parameter	Description
host	Host of the TorchServe server.
port	Port of the TorchServe server (gRPC endpoint).
token	Authentication token for the inference endpoint (optional)
model	Name of a deployed model.
threshold	Confidence threshold for labels (for TSImageLabel only).

For now, the PR is a draft to allow for discussion and optimisation.

[ppanopticon]

The retrieval side of things is currently missing.

Main reason is, that I'm not quite sure how we would use a LabelDescriptor for querying. Is there any existing facility, that does that? I couldn't find any. From where I stand, LabelDescriptors could be used both for classical boolean retrieval.

[ppanopticon]

Added a (preliminary) version of the retrieval side and addressed @lucaro's comments.