Fix SpectralConvS2 bias with channel projection

[fallintoplace]

Summary

• Shape spectral_bias by input channels in the serial and distributed spectral convolution layers.
• Add regression coverage for SpectralConvS2(bias=True) when input and output channel counts differ.

Why

The spectral bias is applied before the grouped channel contraction, while the tensor is still shaped by input channels. It was previously initialized with out_channels, so bias=True failed for channel projections such as 4 -> 8 or 8 -> 4.

Keeping the parameter input-channel-shaped preserves the existing pre-contraction bias semantics and keeps equal-channel cases unchanged.

Validation

• uv run --no-project --with 'torch>=2.6.0' --with 'numpy>=1.22.4' --with parameterized --with pytest python -m pytest tests/test_spectral_convolution.py -q
• uv run --no-project --with ruff ruff check torch_harmonics/spectral_convolution.py torch_harmonics/distributed/distributed_spectral_convolution.py tests/test_spectral_convolution.py

[azrael417]

LGTM, but please amend the Changelog as well here. Thanks for fixing this.

[azrael417]

LGTM.