Fix duplicate-sequence multimer mmCIF chain coordinates

[taivu1998]

Summary

Fixes #525.

This PR fixes DataPipelineMultimer.process_mmcif for multimer mmCIF examples
that contain multiple chains with the same sequence. The pipeline already
deduplicates expensive per-sequence feature generation by caching features under
the sequence string, but the mmCIF path was caching a dictionary after
chain-specific structural labels had been added. Duplicate-sequence chains could
therefore reuse the first chain's all_atom_positions and all_atom_mask.

Root Cause

process_mmcif mixed two feature classes in the same sequence cache:

• cacheable sequence-derived features from the monomer pipeline;
• per-chain mmCIF labels from get_mmcif_features.

When a later chain had the same sequence, it deep-copied the cached dictionary
and skipped get_mmcif_features for its own chain ID.

Changes

• Keep duplicate-sequence caching for sequence-derived features.
• Add a second pass that attaches mmCIF structural features once per actual
  chain.
• Refresh auth_chain_id on cache hits so copied intermediate features remain
  internally consistent.
• Add local shape guards for all_atom_positions and all_atom_mask lengths.
• Add a focused regression test proving:
  • duplicate sequences still run _process_single_chain only once;
  • get_mmcif_features runs for both chains;
  • duplicate-sequence chains receive distinct positions and masks.

Validation

Passed:

python3 -m compileall openfold/data/data_pipeline.py tests/test_data_pipeline_multimer.py
PYTHONPATH=/private/tmp/openfold-issue-525-test-deps-py310 python3 -m unittest tests.test_data_pipeline_multimer
git diff --check

Also passed a real fixture smoke test on tests/test_data/mmcifs/2q2k.cif,
confirming duplicate-sequence chains A and B keep distinct coordinates in
the merged multimer output.

Attempted broader existing tests, but this local environment is missing optional
runtime pieces unrelated to this patch:

• tests.test_data_pipeline: missing attn_core_inplace_cuda
• tests.test_multimer_datamodule: missing pytorch_lightning