Ddsp

Support for training and exporting DDSP-VST models.
Fixes to build.

Major fixes to loudness/power calculations, and standardized padding (center/same/valid). Dataset loaders allow for centered datasets.
Breaks loading/training on older models (twice --> v3), but important to get straightened out before VST release.

What's Changed
* Inference models for VST by copybara-service in https://github.com/magenta/ddsp/pull/403
* Add UrmpMidiUnbatched for unsegmented tfrecords in ddsp.training.data. by lukewys in https://github.com/magenta/ddsp/pull/404
* Fix aliasing in streaming synthesis. by copybara-service in https://github.com/magenta/ddsp/pull/408
* Add Viterbi decoding to PretrainedCREPE, allows hmm decoding during training. by copybara-service in https://github.com/magenta/ddsp/pull/406


**Full Changelog**: https://github.com/magenta/ddsp/compare/v1.7.0...v1.9.0

* MultiDatasetProvider
* Default upload() to mix down to mono for timbre_transfer
* OnlinePreprocessor
* Small changes for compatibility with MIDI-DDSP

* Replace "master" branch with "main"
* Move dataset statistics from colab_utils to postprocessing. ***NB: dataset statistics files will need be regenerated for any models trained before this release.***
* Several small changes for the VST plugin (named outputs of saved model, adding metadata to tflite models, gin config for VST).

Updated SavedModels interface. Scripts for conversion of SavedModels to TFLite and TFJS. Updates to data preparation pipeline to handle very long input files, and enable creating train / eval splits.