Allow changing the back-end for NER filters

[MatheusRich]

An alternative to the current ML approach could be using a local-running (or an open source LLM hosted on a server you own) to identify information like people names and locations. LLMs are particularly helpful because they're flexible (even if faulty) and a simple prompt is enough to "train it" on a new task.

I propose that we allow changing the back-end for NER filters. Maybe instead of passing a model_path, we could build the NER model in the configuration file:

TopSecret.configure do |config|
  config.ner_model = Mitie::NER.new("path/to/ner_model.dat")

  # or use an llm
  config.ner_model = LLMBasedNER.new
end

and LLMBasedNER could implement the required API and do it's work whatever way it wants to. This also has the benefit of only building the NER model once and paying that cost during initialization.

---

Alternatively, this could be achieve by disabling NER filters and providing a new kind of custom filter. The filter interface should be:

1. a #call method that filters an input and returns a
2. a #label method to build the mapping