Add standalone prediction and evaluation references

[wzh20188]

What changed

• add a standalone OpenAI-compatible prediction script for Cypher, GQL, and SQL
• add reference execution-accuracy evaluators for Spanner GQL and TuGraph Cypher
• add a reference grammar, similarity, and Google BLEU wrapper
• add setup, input-format, security, and usage documentation

Why

These scripts make the experiment workflow easier to reproduce while keeping environment-specific evaluation implementations clearly separated as references.

Security and portability

• API keys and database passwords are read from environment variables or runtime arguments
• cloud resource identifiers are supplied at runtime
• machine-specific absolute paths were removed
• generated data, credentials, and service-account files are not included

Validation

• compiled all four Python scripts with python -m py_compile
• verified --help for every script
• verified the repository-relative grammar/similarity evaluator path
• scanned the staged contribution for credential patterns and local absolute paths
• ran git diff --cached --check