Refactor MapDiff into a complete reusable pipeline

[wenruifan]

Summary

• replace the lightweight and duplicated MapDiff example implementation with the complete release-compatible model and diffusion pipeline
• keep data preparation responsible for residue graph and geometry construction while making the collator batch-only
• expose the named mapping workflow used throughout KaleProtein: loader -> preprocessor -> model.embed(**inputs) -> model.generate/predict(**embeddings)
• move reusable sparse EGNN, invariant point attention, and backbone geometry components into kaleprotein
• consolidate IPA pretraining and diffusion training into one real train.py workflow
• update generation, evaluation, documentation, package extras, and smoke coverage

Why

The previous example mixed preprocessing with collation, carried toy and upstream-compatibility copies, and did not faithfully represent the published MapDiff architecture or checkpoint. This made the example harder to understand, reuse, and validate through the Auto APIs.

User impact

The MapDiff example now provides a clean but complete inverse-folding pipeline without importing the original repository. Official CATH processed graph features are preserved, including five-atom coordinates and the real C-beta position. The official v1.0.1 checkpoint is loaded strictly against a checked release-state manifest instead of silently falling back to a partial model.

Validation

• pytest -q: 85 passed
• focused MapDiff and preprocessor tests: 9 passed
• strict load of all 464 tensors from the official MapDiff v1.0.1 checkpoint
• end-to-end generation with the official checkpoint
• Python compile check and git diff --check
• clean wheel build and isolated installation smoke test