Harden provider URLs, response parsing, and tool execution

[luckenco]

Summary

• add explicit tool execution guardrails for per-turn call limits, concurrency limits, and tool timeouts
• run generated synchronous #[tool] functions on Tokio's blocking pool so configured timeouts can still fire
• reject insecure provider base URLs by default, with explicit opt-in helpers for local HTTP endpoints
• harden OpenAI Responses and completion parsing with tagged deserialization, unknown-field rejection, complete output aggregation, redacted tool-argument parse errors, and function-call round-trip coverage
• add dependency-audit CI with a committed Cargo.lock
• update audited transitive dependencies (bytes, rustls-webpki) so the new audit job passes
• reduce duplicated builder/provider and macro naming logic, trim unused dependencies, and document repo/MSRV expectations

Review fixes

• document why FunctionToolCall.r#type needs a serde default
• replace the misleading deserialization test with focused variant-selection and serialization tests
• preserve sequential vs parallel tool-call behavior while enforcing the new local guard limits
• document partial-failure semantics for parallel tool execution
• keep generated sync tool execution timeout-friendly, including context-backed tools

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -Dwarnings
• cargo test --all --verbose
• cargo test --all --examples --verbose
• cargo audit

Notes

• Example binaries were compiled through the example test harness; live example execution requires provider API keys.