feat: transcribe QQ voice messages to text via GLM-ASR

[hilr]

Summary

• Add voice-to-text for QQ channel audio attachments using GLM-ASR-2512 (ZhipuAI). When GLM_API_KEY is configured, audio attachments are transcribed in the channel layer before being forwarded to the router/agent, so the agent receives the spoken content ("用户发送了一段语音，内容: {text}") instead of just a file path.
• When GLM_API_KEY is empty, behavior is unchanged — voice attachments fall back to the file-path-only description.
• New src/agent_box/asr/glm.py client (reuses existing httpx, no new deps). Transcription failures degrade gracefully to None (rendered as the file path), never blocking the message flow.

Test Plan

• tests/test_asr.py — 6 tests: success (request shape verified), HTTP error, empty text, network error, missing file, empty file
• tests/test_channels.py — 8 tests: _atts_text with/without transcription, mixed attachments, _download_atts transcribes audio / skips when ASR disabled / falls back on failure / ignores non-audio, _asr enable/disable based on key
• uv run pytest tests/test_asr.py tests/test_channels.py — 42 passed
• uv run ruff check on changed files — clean

Closes #56