refactor: deduplicate dual-timestep embedding interpolation

[KakaruHayate]

Summary

Review-only follow-up for openvpi/DiffSinger:dual-timestep. This PR keeps the feature behavior unchanged and centralizes the repeated timestep embedding and masked interpolation logic shared by WaveNet, LYNXNet, and LYNXNet2.

Base and scope

• Base branch: openvpi/DiffSinger:dual-timestep
• Base commit: 72981e4
• This PR contains exactly one review-fix commit and changes only four files.

Changes

• Add interpolate_dual_timestep_embedding in modules/commons/common_layers.py.
• Use the helper in WaveNet, LYNXNet, and LYNXNet2.
• Preserve WaveNet post-embedding MLP behavior and each backbone tensor layout.
• Raise a clear error when a mask is passed without the second timestep.
• Remove trailing whitespace introduced by the source branch.

Review source

Implements the CodeRabbit review feedback from KakaruHayate#53.

Validation

• Python compileall: passed for modules/.
• git diff --check: passed.
• Runtime tensor tests were not run locally because the isolated Python environment does not include PyTorch.