Add pinned KRONOS2 native multiplex embeddings

[CraigMyles]

Summary

• add kronos2 as a native multiplex patch encoder using the immutable MahmoodLab/KRONOS2 Hugging Face revision and SHA-verified weights and marker metadata
• support ordered native marker selection and optional novel-marker registration through the upstream BioLinkBERT path, with the conditional text model pinned and every required artifact verified
• bind the effective panel, physical marker indices, normalization, forward contract, and optional registration recipe into output fingerprints and store provenance
• prepare the focused v0.2.1 model update without changing the v0.1 store layout, geometry-only grid_hash, or ordinary brightfield execution

Model loading and execution

• download only the pinned KRONOS2 files required for inference into a
  raw2features-managed cache; reject unexpected files before executing the model's
  custom code
• construct the model entirely from that local pinned copy, then remove the temporary
  Python import path added by the upstream loader
• use the published fp32 inference path and fixed batches of eight; short final batches
  are temporarily padded by repeating their last patch, with the extra outputs
  discarded
• accept square patch sizes divisible by 16, defaulting to 256 px; reject
  torch.compile because compiled execution has not been validated against the
  reference implementation
• use the 288 released marker entries without downloading BioLinkBERT; the additional
  text model is downloaded only when a user explicitly registers novel markers
• do not silently treat Hoechst or DNA1/DNA2 as DAPI because KRONOS2 applies
  marker-specific normalization; excluded nuclear channels remain available for
  segmentation, while explicit novel-marker registration is supported
• keep native marker-aware models such as KRONOS2 separate from RGB encoders within a
  single command; compatible runs can still add their outputs to the same store

Validation

• 678 local non-slow, non-network tests passed; 105 dependency-gated tests skipped and 68 slow/network tests deselected
• the loopback authenticated-HTTP reader test passed separately outside the filesystem sandbox
• Ruff, git diff --check, uv lock --check, wheel/sdist builds, and Twine metadata checks passed
• real-weight L40S validation used Python 3.12, Torch 2.6.0+cu124, Transformers 4.56.0, timm 1.0.19, and xFormers 0.0.29.post3
• verified the pinned KRONOS2 weight and marker-table hashes before loading; the
  raw2features-managed cache contained only its seven required model entries plus
  Hugging Face's cache metadata, and xFormers was active
• on a real uint16 eight-channel IMC OME-Zarr, the published-marker path produced finite deterministic (9, 768) fp32 embeddings and matched the upstream forward bit-for-bit for the first full batch
• the optional pinned BioLinkBERT path registered a novel marker and produced finite (2, 768) fp32 embeddings
• marker reorder, exclusion, and novel registration produced distinct output fingerprints
• an end-to-end CLI run wrote 16 finite (768,) KRONOS2 embeddings with the expected marker-sensitive fingerprint