Support saving the extracted features to disk

I suggest that create a way to save (serialize) the extracted features to disk, and then load it and do the matching directly from there. It is useful in a few cases:

1. Create unit test for the feature extractor, e.g., the binja extractor
2. Separate the feature extractor and the matching process, e.g., for TTD, we might want to run some C++ code to do the feature extraction, save it, and then do the matching elsewhere
3. Write the binja extractor in C++ which is more performant

Start by tracing the binja feature extractor and the matching process to understand the extracted feature representation and their current coupling. Define a serialized format that can be saved, loaded, and consumed by matching, including the cross-language C++ use case. Done means extracted features can be persisted and later used for matching, with coverage for the proposed unit-test workflow.

Written by the indexing model from the issue text.