Fix time-domain convolution kernel orientation - #3083

What does this PR do?

Fixes #1527.

convolve1d(..., use_fft=False) currently performs cross-correlation because it passes the kernel directly to PyTorch's conv1d. For a waveform [1, 2, 3, 4] and kernel [1, 2, 4], it returns [17, 24] instead of the valid convolution [11, 18]. Reverse the kernel along its time axis before calling PyTorch.

This intentionally changes time-domain results for asymmetric kernels. DropFreq uses this path both to build and apply its filter; its existing frequency-rejection tests pass, and old/new waveform comparisons agree within floating-point tolerance (bitwise reproducibility is not promised). The FFT path, including the default reverberate implementation, is unchanged.

Validation on CPU with PyTorch 2.6.0:

• New tests compare against NumPy convolution across padding, stride, dtype, batch and grouped-channel cases, check FFT agreement with linear-convolution padding, and run autograd gradcheck. Before the fix: 24 failed, 1 passed.
• After the fix: the new tests, existing signal-processing and augmentation tests, and signal-processing doctests all pass: 46 passed.
• pre-commit run -a and git diff --check pass. No model or dataset downloads were needed. The full repository test suite and GPU execution were not run.

Implemented and tested autonomously with OpenAI Codex, with a separate agent reviewing the code, callers, and compatibility checks.

PR review