Add native score-only confidence CLI for existing structures

[taivu1998]

Summary

Adds a native score-only workflow for Protenix so users can evaluate existing PDB/mmCIF coordinates with the model trunk and confidence heads without running diffusion sampling.

This addresses #227.

Motivation

Issue #227 asks whether Protenix can be used in a scoring-only mode, especially for complexes. The existing inference path always samples new coordinates through diffusion, while the confidence head already accepts predicted coordinates. This PR exposes that capability through a narrow, native workflow instead of requiring users to rely on a pinned external fork.

Changes

• Adds fixed-coordinate inference support in Protenix.forward() and _main_inference_loop() via score_coordinates.
• Adds InferenceRunner.score() to run the trunk, distogram, confidence head, and existing confidence summary code on supplied atom coordinates.
• Adds protenix score CLI support for PDB/mmCIF inputs and directories.
• Adds structure parsing and coordinate remapping utilities that:
  • preserve source chain IDs when possible through generated JSON id fields,
  • map source coordinates into Protenix featurized atom order,
  • detect duplicate atom keys and missing atoms,
  • support configurable missing-atom fallback policies.
• Writes score outputs:
  • summary.csv,
  • <sample>/summary_confidence.json,
  • <sample>/chain_id_map.json,
  • optional <sample>/full_confidence.json,
  • optional <sample>/scored.cif,
  • failure records when applicable.
• Documents protenix score in the README and training/inference instructions.
• Adds focused unit tests for structure file collection, chain ID preservation, chain mapping, and coordinate remapping.

Design Notes

The implementation deliberately keeps the first version small:

• It reuses the existing inference preprocessing path for MSA, template, and RNA MSA handling.
• It reuses the existing confidence summary implementation instead of adding separate scoring metrics.
• It skips diffusion only when fixed coordinates are explicitly provided.
• It does not add role-aware MSA caches, target/binder-specific outputs, ipSAE-style metrics, or multi-model scoring in this PR.

Validation

Ran:

git diff --cached --check
uvx ruff check protenix/model/protenix.py runner/inference.py protenix/data/inference/structure_scoring.py runner/scoring.py runner/batch_inference.py tests/test_structure_scoring.py
python3 -m py_compile protenix/model/protenix.py runner/inference.py protenix/data/inference/structure_scoring.py runner/scoring.py runner/batch_inference.py tests/test_structure_scoring.py
uv run --python 3.11 ... python -m pytest tests/test_structure_scoring.py

Results:

• Ruff passed.
• Python compilation passed.
• tests/test_structure_scoring.py passed: 7 tests.
• Verified the Click group exposes score --help with LAYERNORM_TYPE=torch.

Local Environment Notes

I did not run a full model-backed protenix score on an example structure locally because this machine is missing the Protenix CCD runtime data file at /Users/vuductai/common/components.cif, which blocks the repo's normal parser path. On this macOS host, CLI import smoke also requires LAYERNORM_TYPE=torch to avoid the existing CUDA fused-layernorm compile path.