Remove Faker fill and preserve strict PARSE semantics

[skadel]

Summary

• remove Faker-based filling so all generated values stay coherent with the LLM-authored scenario
• remove the unused Faker dependency and obsolete regression coverage
• preserve strict BigQuery PARSE_* behavior while keeping SAFE.PARSE_* tolerant
• update sqlglot 30.12 canaries and scalar-folding expectations

Validation

• poetry run pytest -q --disable-warnings — 2095 passed, 1 xfailed
• focused parsing/generator tests — 80 passed
• Ruff checks passed
• git diff --check passed