fix(llm): normalize OpenAI strict schemas

[haasonsaas]

Summary

• Normalize OpenAI strict JSON schemas before json_schema response_format or strict tool-call transport.
• Seal every object, require all remaining properties, strip unsupported validation keywords, fold trivial allOf refs, and rewrite definitions refs to $defs refs.
• Inline direct property refs inside anyOf variants because the live OpenAI API rejects the browser action union when find_elements is only a nested $defs ref.

Testing

• cargo test -p browser-use-llm openai_json_schema_mode_makes_schema_strict -- --nocapture
• cargo test -p browser-use-llm
• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• python3 scripts/release-version.py --check
• python3 scripts/release-version.py --self-test
• python3 scripts/upstream-drift.py --self-test
• python3 scripts/upstream-drift.py
• git -c core.fsmonitor=false diff --check
• MCP protocol smoke: initialize, ping, tools/list, browser_use_state
• MCP OpenAI default browser_use_agent with gpt-4.1-mini: completed in one step with exact done text and all_errors: []
• MCP OpenAI prompt-only agent smoke: completed in one step with exact done text and all_errors: []
• MCP OpenAI tool-call agent smoke: completed in one step with exact done text and all_errors: []

Release-Impact: patch