[Enhancement] Expose inference speed metrics for Chat and Completion

[cch1rag]

Summary

Closes #315.

This exposes the backend timing data on the Unity side.

After a completed Chat() or Completion() call, you can now read:

• TokensPerSecond
• PromptTokensPerSecond

These values come from the timing data already returned by the backend. If timing data is not available, both properties stay at -1.

What changed

• added TokensPerSecond and PromptTokensPerSecond to LLMClient
• updated the managed wrapper to request the JSON completion response when needed
• parsed the response JSON and still returned plain text to callers
• used the same parsing path for both Completion() and LLMAgent.Chat()
• added tests for JSON parsing, fallback behavior, and stale timing reset

Testing

• git diff --check
• Unity EditMode parser tests passed:
  • LLMUnityTests.TestLLMClient_ResponseParsing

I kept the README update and the package test-discovery cleanup out of this PR to keep it focused on the issue itself. I can send those too if they would be useful.