Optimize CPU featurization pipeline: 37% speedup

[longleo17]

This PR accelerates the data pipeline, a CPU bottleneck in current workflows

Summary

• Featurizer encoder: Replace manual one-hot dictionary construction with torch.nn.functional.one_hot, eliminating per-call dict comprehension overhead
• Atom name encoding: Vectorize ref_atom_name_chars_encoded using numpy.frombuffer on ASCII bytes + F.one_hot, replacing nested Python loops over characters
• Template featurizer: Pre-allocate output numpy arrays instead of list append + np.stack; reuse a shared DistogramFeaturesConfig instance across templates
• Template parser/utils: Batch numpy operations for coordinate extraction, atom mask computation, and position lookups
• Dataset filtering: Replace df.apply(lambda x: x in set) with vectorized df.isin() for eval_type filtering

Benchmark results

Measured end-to-end on representative PDB complexes with template+MSA workloads:

• ~37% wall-clock reduction in CPU featurization time
• No changes to model outputs (numerically identical features)

Files changed

File	Change
protenix/data/core/featurizer.py	Vectorized one-hot encoding + atom name encoding
protenix/data/template/template_featurizer.py	Pre-allocated arrays, shared config
protenix/data/template/template_parser.py	Vectorized coordinate extraction
protenix/data/template/template_utils.py	Batch numpy operations
protenix/data/pipeline/dataset.py	Vectorized pandas filtering