fix: honor short voice splitter chunks

[he-yufeng]

Summary

• send non-empty custom splitter chunks to TTS even when they are shorter than 20 characters
• preserve turn_ended when a custom splitter consumes the whole buffer before _turn_done()
• add a regression test for short custom chunks

Fixes #3363.

To verify

• uv run pytest tests\voice\test_pipeline.py -q --basetemp .tmp\pytest-voice
• uv run pyright src\agents\voice\result.py tests\voice\test_pipeline.py
• uv run ruff format --check src\agents\voice\result.py tests\voice\test_pipeline.py
• uv run ruff check src\agents\voice\result.py tests\voice\test_pipeline.py
• uv run python -m py_compile src\agents\voice\result.py tests\voice\test_pipeline.py
• git diff --check

[seratch]

Resolved by #3364