fix(wizard): move LLM test into Rust core & add no_reasoning_control option

[erning]

Summary

Fixes #56 — LLM correction silently falls back to raw ASR text when the LLM call times out, but the wizard's Test Connection passes because it uses a completely different code path.

Root cause: The wizard sent a minimal "Hi" message via Obj-C with a hardcoded 15s timeout, while the runtime sends full system/user prompts through Rust with a default 8s timeout. Reasoning models like GLM-5-turbo take 15+ seconds (ignoring reasoning_effort: "none"), so the runtime always timed out while the test always passed.

Changes

• Move LLM test into Rust core — sp_llm_test() FFI function that shares the exact same correct() code path as runtime: same prompts, dictionary, timeout, temperature, top_p. The Obj-C wizard just calls this and displays the JSON result.
• Always report elapsed time — test_correction() returns (Result<String>, Duration) so elapsed time is available even on timeout/error (e.g., "LLM correction timed out (8.0s)" instead of "(0.0s)").
• Add no_reasoning_control config option — New LlmNoReasoningControl enum:
  • reasoning_effort (default): sends "reasoning_effort": "none" — works for OpenAI o-series
  • thinking: sends "thinking": {"type": "disabled"} — works for GLM and similar models
  • none: sends nothing

Test plan

• cargo build succeeds
• Xcode build succeeds
• Test Connection with valid config shows "Connection successful! (X.Xs)"
• Test Connection with timeout_ms: 1000 shows timeout error with correct elapsed time
• Setting no_reasoning_control: thinking with GLM-5-turbo responds in ~2s instead of 15s+
• Setting no_reasoning_control: none sends no reasoning parameters
• Hold-to-talk and tap-to-toggle still work