feat: add real streaming to agentic tool loop (phase 2)

[pavelanni]

Summary

Add real token-by-token streaming to the agentic tool loop
(pkg/tools/loop.go) so phase 2 responses stream from the provider
instead of using simulated word-by-word chunking.

Context

Issue #44 adds StreamWithTools to the provider interface. The first
iteration (PR TBD) implements real streaming for phase 1 (no tools)
and keeps simulated streaming for phase 2 (tool loop). This issue
tracks the follow-up to make phase 2 streaming real.

What needs to change

• RunToolLoop accepts a streaming callback parameter
• On the final iteration (when StopReason == EndTurn), use
  StreamWithTools instead of CompleteWithTools and pipe text
  deltas through the callback
• All consumers benefit: OpenAI API, Chat TUI, A2A server

Depends on

• feat: add streaming responses to LLM provider interface #44 (provider streaming interface)