feat(bookstack-agent): stream tokens in real time instead of buffering

[amrit110]

Problem

The BookStack agent response was loading in one shot — the user saw nothing until the entire LLM response was ready, then everything appeared at once.

Root cause: ask_stream() was silently consuming the entire Anthropic streaming response into memory (accumulating all tokens, waiting for get_final_message()), and only then emitting events. The character-by-character loop at the end was fast synchronous Python iteration over an already-complete string, so all text_chunk events flushed together in a single TCP write.

Fix

• src/aieng_bot/bookstack/agent.py: Yield text_chunk events as tokens arrive from the streaming API instead of buffering. When a content_block_start for a tool_use block is detected and text was already streamed (reasoning/planning text from the model), emit text_clear to tell the UI to discard it. The answer event is emitted at the end as the authoritative complete text.
• tests/bookstack/test_agent.py: Add _make_tool_use_block_start_event helper and a new test test_stream_text_clear_emitted_when_text_precedes_tool_call covering the discard path.

The UI already handled both text_chunk and text_clear events — no frontend changes needed.

Test plan

• All 12 unit tests pass (pytest tests/bookstack/test_agent.py)
• Manually verify the chat UI streams the response token-by-token
• Verify tool steps still appear correctly before the final answer
• Verify reasoning/thinking text is discarded (shows text_clear) when the model calls a tool