This commit adds support for using inter-chain template information when
scoring protein complexes, enabling AF2Rank-style evaluation.
Changes:
- Add allow_interchain_templates parameter to Diffuser, Evoformer, and
 _embed_template_pair in model.py
- When True, template embeddings include both intra-chain and inter-chain
 distances. When False (default), only intra-chain template features are
 used (standard AF3 behavior)
- Add --allow_interchain_templates CLI flag to run_af3score.py
- Thread parameter through full call chain: CLI → process_fold_input →
 predict_structure → run_inference → model
Use case:
For scoring existing structures (--init_guess=true --diffusion_steps=0),
this allows the template structure's full complex geometry (including
interface contacts) to inform confidence predictions.
Backward compatible: defaults to False to preserve standard AF3 behavior.

- Add command-line argument parsing to all 4 example scripts
- Scripts now accept -i/--input_dir and -o/--output_dir flags
- Remove hardcoded paths (./pdb, ./complex_chain_cifs, etc.)
- Enables flexible directory structure for pipeline scripts
- 1_extract_chains.py: Supports both PDB and CIF input files
- 3_generate_json.py: Added -c/--csv_path flag for CSV input
- 4_extract_iptm-ipae-pae-interaction.py: Modified update_sc_file() to accept parameters

- 1_extract_chains.py: Save complex_chain_sequences.csv to output directory instead of working directory
- 4_extract_iptm-ipae-pae-interaction.py: Save failed_records.txt to same directory as output file