Fix output amplitude scaling in iterative Wiener filter (#420)

[deekshaNVIDIA]

Summary

Fixes #420.

IterativeWiener.compute_filtered_output divided the frame DFT by sqrt(frame_len) before applying the Wiener gain and returning the spectrum:

frame_dft /= np.sqrt(self.frame_len)
...
return self.wiener_filt * frame_dft

The spectrum is then handed to STFT.synthesis, which uses an unnormalized rfft/irfft pair (pyroomacoustics/transform/dft.py). So the returned spectrum must be on the same scale as stft.X. The extra 1 / sqrt(frame_len) factor therefore scaled the enhanced signal down by sqrt(nfft) (and, as a side effect, mutated the caller's stft.X in place).

This matches the report: rescaling the output by sqrt(nfft) recovered the input amplitude for nfft = 512, 1024, 2048.

Root cause / why removing the line is correct

The divided frame_dft was only used in two places:

1. The returned spectrum self.wiener_filt * frame_dft — this is the bug.
2. The intermediate frame s_i = irfft(self.wiener_filt * frame_dft), which feeds lpc(s_i, ...).

LPC coefficients are scale invariant (they depend on the normalized autocorrelation), and the gain g^2 is computed separately from current_frame (not from s_i). So the Wiener gain values are unchanged whether or not the division is applied — only the output amplitude changes. Removing the division fixes the amplitude and also removes the in-place mutation of stft.X.

Verification

Before (installed 0.10.1), output/input RMS ratio for a near-passthrough harmonic signal:

nfft	ratio	ratio × √nfft
512	0.0442	1.000
1024	0.0312	1.000
2048	0.0221	1.000

After the fix the ratio is ~1.0 for all sizes, and the relative denoising behaviour is unchanged.

Changes

• Remove the erroneous frame_dft /= np.sqrt(self.frame_len) scaling in compute_filtered_output.
• Add test_iterative_wiener_output_scaling, a regression test asserting output RMS ≈ input RMS across several FFT sizes (fails on the old code with ratio ≈ 0.044, passes now).
• Add a CHANGELOG entry.

Test plan

• pytest tests/denoise/test_iterative_wiener.py passes (2 passed).
• New test fails on the previous (buggy) implementation.
• black --check and isort --check-only pass on the changed files.