Fix Piscis legacy model loading with converted checkpoint

[arjunrajlab]

Summary

• pin the Piscis image build to the converted rajlab/ps_20240419_112256 PyTorch checkpoint
• require the .pt filename instead of downloading an extensionless legacy JAX checkpoint
• make download_models.py testable while preserving the built-in Piscis model preload
• add focused regression coverage and document the runtime-ready checkpoint requirement

The converted checkpoint is published at https://huggingface.co/rajlab/ps_20240419_112256 and is the sole model in the Raj Lab Piscis collection.

Root cause

The legacy ps_20240419_112256 repository exposed an extensionless JAX checkpoint. Piscis therefore attempted an on-the-fly JAX-to-PyTorch conversion at worker runtime, which imports Flax. After the worker image was slimmed and Flax was removed, selecting that model failed with ImportError: Flax is required for converting JAX models.

Publishing and downloading a pre-converted .pt checkpoint keeps the inference image torch-only and avoids runtime conversion.

Validation

• .cache/testvenv/bin/python -m pytest workers/annotations/piscis/tests -q — 2 passed
• Python compile check for the downloader and regression test
• git diff --check
• live download through the updated downloader
• downloaded SHA-256: f281a239d1d5075ee2342c4772ba7167c441f878e5345c632436caf3b43d320a
• checkpoint structure: 583 tensors plus metadata