Change variance retaking masks from zero inputs to zero embeddings

[yqzhishen]

There are 4 modifications to the variance retaking mechanism in this pull request:

1. Zero masks are applied to embeddings instead of inputs.
2. The retaking embedding for variance parameters is removed.
3. Masks of each variance parameter are decoupled so that they can be retaken separately.
4. Generation algorithm of retaking masks is improved.

NOTE: These modifications are backward-incompatible. All models involving pitch and variance predictions need to be re-trained.