add PathForge-native feature extractors and processor support

[JurreWeijer]

Summary

This PR adds PathForge-native feature extractors and defines how slide processors can support them now and in the future.

PathForge-native extractors

• Added FeatureExtractorBase, the shared interface for PathForge feature extractors:
  • load and own a model;
  • prepare image patches for that model;
  • turn batches of patches into feature vectors;
  • support evaluation mode and moving to a device.
• Added helpers to register, find, and build PathForge-native extractors.
• Added Mascaret and Phaet as PathForge-native extractors, including their preprocessing, output checks, and guidance for accessing their gated Hugging Face weights.

Processor integration and LazySlide support

• Extended SlideProcessorBase so processors can:
  • list the feature extractors they can run themselves;
  • state whether they can also run PathForge-native extractors.
• Added FeatureExtractionRequest to keep extractor selection separate from runtime settings such as batch size, workers, AMP, and device.
• Added the first adapter in LazySlide:
  • LazySlide and timm models are still passed directly to LazySlide as processor-native models;
  • PathForge-native extractors are built when feature extraction starts and wrapped so LazySlide can run them;
  • image patches are converted to the shared RGB uint8 format before PathForge preprocessing is applied.
• If the same name exists both as a LazySlide/timm model and as a PathForge-native extractor, the processor-native model is used. This is logged clearly.

Registry and processor layout cleanup

• FEATURE_EXTRACTORS now contains only PathForge-native extractor classes that PathForge can build itself.
• LazySlide and timm model names are no longer added to this registry. They are provided by the processor and can only be run by that processor.
• Available extractors and configuration validation are now determined from the selected slide-processing processor, without loading models during validation.
• Reorganized slide processing so each processor has its own folder, for example slide_processing/lazyslide/.
• Added shared processor code above these folders, including the new build_slide_processor() builder.
• Updated slide-retrieval code to use the shared builder instead of repeating processor-loading logic.

Verification

• Added tests for registry behaviour, available extractor selection, name collisions, LazySlide support, patch conversion, and Mascaret/Phaet.
• Added a LazySlide smoke test that runs a registered PathForge-native extractor and checks the output features, timing, and memory use.
• Updated documentation for feature-extractor availability and Hugging Face authentication through hf auth login or HF_TOKEN.