[FEATURE] add pytta.psychoacoustics with loudness, sharpness and roughness - #79

Adds pytta.psychoacoustics with three metrics of a one-channel SignalObj in Pa. Each returns a dict of floats and NumPy arrays, with the keys listed in the docstring.

1. loudness_zwicker(signal, field='free', method='stationary', time_skip=0.0): ISO 532-1:2017, stationary and time-varying methods, free or diffuse field. It gives N (sone), LN (phon) and the specific loudness over 240 steps of 0.1 Bark. The time-varying method adds N(t), LN(t) and N'(t) every 2 ms, with Nmax and N5.
2. sharpness_din45692(signal, field='free', method='stationary', time_skip=0.0): DIN 45692:2009, Formula (1), applied to that specific loudness with k = 0.11. It gives S, or S(t) and S50 for the time-varying method.
3. roughness_daniel1997(signal, time_skip=0.0): the model of Daniel and Weber (1997). It gives R(t), R'(t) over 47 half-Bark channels and Rmean (asper).

Regarding the origin of the code, loudness is a port of Loudness_ISO532_1.m from SQAT (https://github.com/ggrecow/SQAT), which is BSD-3-Clause: code by Ella Manor for AARAE, after the C code of the standard. Roughness is a port of Roughness_Daniel1997.m and its utilities, which are MIT: code by J. E. Schrader and Dik Hermes. Both notices are reproduced in the module header. No GPL file of SQAT was used. Sharpness is written from the text of DIN 45692. Signals not at 48 kHz are resampled the way MATLAB resample does, which matches it to 2e-15.

The tests in tests/test_psychoacoustics.py (13 tests) synthesise their signals, so no file of the standards is needed:

1. Parity with SQAT on golden data from tests/golden/make_loudness_sqat.m and make_roughness_sqat.m (MATLAB R2026b, same float32 samples). Stationary loudness agrees within 3e-13 sone, time-varying loudness within 3.1e-8 sone and roughness within 2.8e-5 asper.
2. ISO 532-1 Annex B, test signals 1 to 4. Signal 1, given as band levels, gives 83.2957 sone against 83.296, and the tones 2 to 4 fall within the tolerance of the standard (5 % or 0.1 sone).
3. DIN 45692 Tables A.2 and A.3: eight band noises at 4 sone fall within 5 % or 0.05 acum, as do the reference noise and its S50.
4. Units. A 1 kHz tone at 40 dB gives 0.999976 sone, and the reference AM tone (1 kHz, 60 dB, 70 Hz, m = 1) gives 1.000001 asper.

Outside the test suite, the code was also run on the signal files that come with the standards. ISO 532-1 signals 2 to 5 agree within 0.034 sone (0.26 %), and signals 6 to 25 keep N(t) within 5 % or 0.1 sone at every frame. All 41 signals of DIN 45692 Tables A.2 and A.3 fall within 0.45 of the tolerance, the worst being the 10.5 kHz band at -0.19 acum.

Checked with Python 3.12, numpy 1.26 and scipy 1.17. With NumPy 2 the tests fail in SignalObj on np.complex_, which #53 fixes.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com