[DiffSinger/BUGFIX]Implement DiffSinger variance retake with hard compose

[KakaruHayate]

Merge order and dependencies

1. Merge or rebase Support DiffSinger pitch local retaking #2183 first. It provides the DiffSinger pitch retake frame-mask and non-contiguous retake write-back foundation.
2. Merge this variance retake change next. It replaces the variance crossfade and cumulative baseline behavior introduced by Add DiffSinger variance local pitch patch #2176 with tensor-cache-based partial retake and hard composition.
3. Rebase or merge Refresh real curves after phrase rendering #2175 after this change so the real-curve scheduler consumes the final hard-composed variance result.
4. Process Add prioritized render scheduling #2174 last. Its render-priority changes are ordering-only and should be reconciled with Refresh real curves after phrase rendering #2175 changes in RenderEngine.cs and Notifications.cs.

This PR is intentionally based on current master, where #2176 is already merged. It does not include the code from #2183, #2175, or #2174. The order above describes the intended final integration order; if this PR is reviewed before #2183 is merged, it can remain independent on master and be rebased when the stack is integrated.

Retake scope requirements

Partial variance retake is allowed only when the tensor-cache scope remains compatible:

• Pitch and frame-level speaker embedding changes are retake-eligible.
• Linguistic encoder inputs or outputs, phoneme durations, word durations/divisions, phrase frame layout, model identity, variance channel layout, and diffusion steps/speedup remain part of the scope key. A change to any of them uses a different scope and performs full generation.
• Tensor cache must be enabled and a compatible previous result must exist.
• The first version does not expand the frame mask.
• No crossfade is used.
• Model output outside the mask is not trusted. Hard composition restores every unmasked frame from the previous cached result.

What changed

• Split complete variance cache lookup from the partial-retake scope.
• Reuse cached variance curves as predictor inputs during partial retake.
• Build a frame mask from pitch and speaker-embedding changes, then expand it to the model's [1, T, V] retake input.
• Hard-compose predicted frames with the previous result after inference.
• Save the final composed result to the complete tensor cache so undo/redo can reuse exact historical states.
• Remove the fixed 50 ms crossfade, weighted blending, and cumulative merged baseline.
• Add focused tests for frame masks, frame-grouped speaker embeddings, channel expansion, hard composition, leakage prevention, and incompatible metadata.

Validation

• dotnet restore OpenUtau.Test/OpenUtau.Test.csproj
• dotnet test OpenUtau.Test/OpenUtau.Test.csproj --no-restore --filter "FullyQualifiedName~DiffSingerVariancePatchTest"