Add masked BLEU and RTF result audits

[luojiaxuan]

Summary

• Add the target-term-masked BLEU computation path to the offline StreamLAAL evaluator and a release aggregation script.
• Publish the masked-term BLEU TSVs, RASST-vs-InfiniSST comparison table, artifact map, and PDF/PNG figure under docs/results/main_result_global_cache30_30_20_20/.
• Add the ACL tagged En-Zh InfiniSST validation TSV backfill used by the artifact map and document the audit in the root/result READMEs.
• Publish the RAG compute RTF package under docs/results/rag_compute_rtf/, including frozen TSV data, PDF/PNG figure, plotting script, provenance README, and root README link.

Validation

• python3 -m py_compile code/rasst/eval/offline_sst_eval/offline_streamlaal_eval.py code/rasst/analysis/main_result/compute_masked_terms_quality.py code/rasst/analysis/main_result/plot_masked_terms_bleu.py
• python3 -m py_compile docs/results/rag_compute_rtf/plot_rag_compute_rtf.py
• sha256sum for the new masked-BLEU TSVs/figures and RTF package matches the checksums recorded in the result READMEs.
• masked_terms_quality.tsv contains 48 ok rows, with 4 rows for each dataset/method/language group.
• docs/results/rag_compute_rtf/data.tsv contains LM 1-4 rows with median RAG RTF values 3.8497%, 2.2055%, 1.4778%, and 1.1366%.
• git diff --cached --check passed before the RTF commit; generated TSV line endings were normalized to LF.

Note: raw *.log hypothesis/runtime files remain ignored by repo policy; the uploaded TSVs retain their artifact pointers and computed metrics.