DiffSinger: support SHFC (pitch shift curve)

[yqzhishen]

• In normal procedure, SHFC only shifts the pitch on variance predictors:
  • variance predictor (shifted pitch) => acoustic model (original pitch) => vocoder (original pitch)
  • If the voicebank has no variance model, SHFC is a no-op.
• With pitch controllable vocoders, SHFC shifts pitch for both variance predictors and acoustic models, meaning synthesizing on shifted pitch and then shift back to original pitch with styles and formants unchanged using the vocoder:
  • variance predictor (shifted pitch) => acoustic model (shifted pitch) => vocoder (original pitch)