Better normalization of variance parameters

[yqzhishen]

Motivation

• Absolute values of some variance parameters are two large. Normalize them to the similar scale may improve the performance and benifits fine-tuning.
• Sometimes linear normalization has trade-offs between precision and range. For example, delta_pitch needs more precision around 0, but also needs a wider range than the current default ±8 keys in some situations.

TODO

• New option to normalize variance parameters to (-1, 1) before they are embedded into the model
• Multiple types of normalization: linear, tanh, etc.
• Generalize configuration schemas of all parameters