Use TRIDENT patch encoder factory in HEST benchmark and remove legacy bench model zoo

[guillaumejaume]

This PR migrates HEST benchmark encoder loading to TRIDENT’s patch encoder stack and removes redundant HEST-specific model-loading code.

What changed

• Switched benchmark encoder construction to TRIDENT (encoder_factory) instead of HEST’s local cpath_model_zoo.
• Updated benchmark encoder names/configuration to TRIDENT-native names (e.g. hibou_l, kaiko-vitb8, etc.).
• Removed HEST-local checkpoint registry usage (src/hest/bench/local_ckpts.json) and delegated model weight resolution to TRIDENT.
• Improved benchmark metric readability
• Updated HEST dependency to install TRIDENT with patch-encoder extras (trident[patch-encoders]) so required optional encoder deps (e.g. CONCH/MUSK) are available.
• Pinned transformers to <5 to fix Hibou-L loading incompatibility (transformers.onnx import issue with v5).

@pauldoucet Main todos:

• put trident on pypi as trident-wsi (version 0.2.4 is ready, see PR on trident repo)
• test core functions of HEST (basis tests look ok)
• re-run bench on all trident models (running now on my end)

[pauldoucet]

Hi @guillaumejaume, thanks for the changes :) !

I'll test on my end ASAP, probably this weekend. There are quite a lot changes, it's probable that something will break

[guillaumejaume]

@pauldoucet, I re-ran the benchmark on all trident patch encoders, results are visible in the updated README. Custom patch encoders are also working fine. Segmentation and patching seem fine. Tutorials run ok. But there are small things that I don't know.

Feel free to lmk what should be tested.

[pauldoucet]

looks good to me! If the tutorials already run, that's a great start

I think to be a 100% sure, maybe trying to process a Xenium or Visium sample from start to finish and compare with existing samples on HF could help (section Adding new samples to HEST)

[pauldoucet]

I think the previous workaround worked well, have you encountered another issue here?

[pauldoucet]

@guillaumejaume thanks a lot for the changes! I also tried processing a mock Xenium sample, seems like it's working fine.

For some reason, the segmentation CI test is hanging, I just commented it out for now.

Feel free to merge if you've already pushed all your changes