Fix normalization of integer and silent signals

[gaoflow]

Summary

• promote integer-valued signals before peak normalization so NumPy can represent fractional results
• preserve zero-amplitude signals instead of dividing by zero and producing nan
• cover dtype limits, bit scaling, and normalized silent WAV output with regression tests

Problem

normalize() currently performs division in place on a copy of the input. Integer arrays therefore raise UFuncOutputCastingError, while a silent floating-point signal is divided by a zero peak and becomes nan. The latter can propagate into files written by MicrophoneArray.to_wav(norm=True).

Testing

• 571 passed with the full test suite
• black --check pyroomacoustics/utilities.py tests/test_normalize.py
• isort --check-only --diff --profile black pyroomacoustics/utilities.py tests/test_normalize.py

[gaoflow]

The lint check failed before installing or running any linter. actions/setup-python@v4 did not find Python 3.9 in the runner cache, attempted to download it, and received GitHub's Unicorn HTML error page. Local checks pass with the workflow versions (black 26.1.0 and isort 7.0.0), and the full suite passes (571 tests). Fork authors cannot rerun an upstream workflow job, so I am documenting the infrastructure failure here: https://github.com/LCAV/pyroomacoustics/actions/runs/29541362105/job/87764109869

[fakufaku]

Thanks for the improvement!

[fakufaku]

Python 3.9 has been deprecated 8 months ago, so it might be time to drop it. Thanks for documenting!