[pytx] New "Classifier" Core Interface

[Dcallies]

Ref: https://github.com/facebook/ThreatExchange/blob/main/python-threatexchange/ROADMAP.md

pytx was built with hashing and matching first, but many companies need help adopting services that don't share embeddings directly, such as Google's content safety API, or Amazon Rekognition, or even a self-trained image or video classifier.

We can support these with a new core interface alonside SignalType, ContentType, and ExchangeAPI: Classifier / ClassifierType (to your preference).

The entirety of the core pytx offering, taking an image -> hashing it -> comparing it to known indexed hashes -> returning "Match" or "No Match" is actually itself a type of classifier, so the first implementation might just be bunding that service. Considering use the the matching core interface from HMA as inspiration.

Interfaces tend to be the hardest thing to get right, and tend to be hard to change, so this will involve a lot of discussion.

[Dcallies]

Some quick ideas, riffing on https://github.com/facebook/ThreatExchange/blob/main/python-threatexchange/threatexchange/signal_type/signal_base.py

class ClassifierType:
    # A human friendly name (I can't decide if this is better than using a function)
    NAME: str

    @classmethod
    @abc.abstractmethod
    def get_with_default_settings(cls) -> self:
        """
        Get an instance of the classifier, configured, authenticated, and ready to use.

        Throws an exception if it fails to initialize
         
        potentially configuring itself with settings files or environment variables."""
        pass

    @classmethod
    @abc.abstractmethod
    def get_content_types(cls) -> t.List[t.Type[content_base.ContentType]]:
        """Which content types this Classifier applies to (usually just one)"""
        pass

    @classmethod
    @abc.abstractmethod
    def classify(self, content) -> ClassificationResult:
        """Run the classifier, which may need to communicate over network"""
        pass

class ClassificationResult:
   
    def __str__(self) -> str:
       """
       A human friendly representation of the result, for display on UIs and logs.

       For example, if your classifier returns a series of labels and scores, this
       could be the json representation, like:
 
         {"nudity": 0.21, "outdoors": 0.01, "dog": 0.99, "cat": 0.61}

    def is_match(self) -> bool:
        """
        A simple flattening of the labels to a bool. 
        
        For many systems this is a shorthand for "is violating/harmful".
       """
       return false  # If you classifier can't be simplified to a bool, return false