Summary

Adds ResponsesWebSocketSession for OpenAI's Responses API WebSocket mode — a stateful session supporting warmup, streaming, multi-turn chaining, and clean shutdown.

Key design decisions

• response.completed as source of truth over response.done — it carries a fully typed CompletionResponse and exists in both SSE and WebSocket codepaths. response.done is treated as a trailing echo, silently absorbed between turns via pending_done_response_id.
• Restricted to Client<reqwest::Client>, non-WASM only — WebSocket transport uses tokio-tungstenite directly, bypassing HttpClientExt. Exposing it on generic H would be misleading since the custom backend gets ignored.
• Reuses StreamingCompletionChunk from the SSE layer rather than duplicating event types.

Other changes

• Fixed ContentPartChunkPart / SummaryPartChunkPart serde rename_all = "snake_case" to match OpenAI's wire format, with deserialization tests
• tokio-tungstenite added as target-scoped native-only dependency

Test plan

• cargo fmt, cargo clippy --all-features --all-targets, cargo test -p rig-core
• Manual test with OPENAI_API_KEY via examples/openai_websocket_mode.rs

Fixes #1425