[codex] Optimize PyPropel hot paths

[Axfff]

Summary

• Vectorize MSA column/alignment frequency counting and MSA one-hot/frequency representations.
• Reduce residue-distance atom-pair loop overhead with NumPy coordinate blocks while preserving existing output behavior.
• Reduce repeated list concatenation in feature assembly paths.
• Lazy-load public PyPropel wrapper modules so plotting dependencies are not imported on plain package import.
• Add focused behavior tests and benchmark scripts for the optimized paths.

Validation

• /Users/zhaoj/Project/TMInter/.venv/bin/python -m unittest discover -s tests
• Performance worktree result: Ran 14 tests ... OK

Benchmarks

Representative local benchmark results from the TMInter environment:

• Alignment columns: 0.300489s -> 0.018815s
• Alignment whole frequency: 0.333040s -> 0.032574s
• MSA one-hot: 0.565160s -> 0.001411s
• MSA frequency matrix: 10.563461s -> 0.012403s
• Distance one2one_all: 0.024435s -> 0.004458s
• Feature assembly solvpred: 0.003100s -> 0.001815s
• Feature assembly blast_: 0.041790s -> 0.009975s

Created as a draft for maintainer review.