fix(voice): honor WAV sample width

[FU-max-boop]

Summary

• encode WAV frames according to the declared 1-, 2-, 3-, or 4-byte sample width
• handle unsigned PCM8 and signed PCM24/PCM32 conversion for both int16 and float32 inputs
• preserve the existing default PCM16 quantization and reject unsupported widths

Problem

VoicePipelineConfig exposes sample_width, and the WAV header already uses it, but frame bytes were always written as int16. For widths other than 2, the header and payload disagree, so readers interpret the audio with the wrong frame boundaries and values.

Validation

• uv run pytest tests/voice/test_input.py -q — 19 passed
• uv run pytest -q — 8170 passed, 33 skipped
• Ruff check and format check
• mypy and Pyright on src
• git diff --check