fix(tokens): add stream options for usage inclusion

[andev0x]

• update openai.go to include stream options
• modify openrouter.go to include stream options
• adjust ui/stream.go to handle usage from stream options

[andev0x]

Fixes

1. internal/providers/openai.go — Two bugs fixed:

• Added streamOptions struct + StreamOptions field on openaiRequest; ExecuteStream now sends stream_options: {"include_usage": true} so OpenRouter emits the final usage chunk
• Moved chunk.Usage capture before the len(chunk.Choices) == 0 guard — the final usage chunk has "choices": [] so it was being silently dropped

2. internal/providers/openrouter.go — Same two bugs:

• Added StreamOptions to openrouterRequest and set it in ExecuteStream
• Moved usage capture before the choices-empty check

3. internal/ui/stream.go — Replaced concrete type assertion rawStream.(*providers.StreamResult) with a usageProvider interface check. Both OpenAIStreamResult and OpenRouterStreamResult implement Usage() (input, output int) but were invisible to the old ollama-only type assertion — so the fallback len(content)/4 (user prompt only, missing system+context) was always used, showing 4-7 tokens instead of ~1000.
4. internal/llm/openai.go — Added streamOptions struct + StreamOptions field to openAIReq; StreamResponse now sends stream_options: {"include_usage": true}.