fix(proxy): strip cache_control from OpenAI conversion, guard tool_choice, handle truncated streams

[thedavidweng]

Summary

• Strip cache_control fields from Anthropic→OpenAI format conversion (system messages, text blocks, tools) to prevent 400 errors on strict OpenAI-compatible endpoints
• Always simplify single text block content to plain string format
• Move tool_choice inside the tools guard so it is omitted when tools is empty
• Handle truncated chat streams with a three-way guard: complete normally when finish_reason is present, emit incomplete response when output exists without finish_reason, emit failed event for empty truncation

Changes

src-tauri/src/proxy/providers/transform.rs

• Remove cache_control passthrough from system message construction, text block construction, and tool construction
• Remove cache_control inheritance logic from normalize_openai_system_messages
• Simplify single text block content to always use plain string format

src-tauri/src/proxy/providers/transform_codex_chat.rs

• Wrap tool_choice assignment inside the if !tools.is_empty() guard

src-tauri/src/proxy/providers/streaming_codex_chat.rs

• Add three-way guard in finalize() for truncated streams: when finish_reason is absent, check if output exists and emit either response.completed with incomplete status or response.failed with stream_truncated error type

Closes #257