Neuralprophet

Included in this release:
* bugfix for Torch 1.9.0 (missing torch.pi)
* New docstrings in Numpy format for most files
* increase training time for better stability
* speed up learning rate range test
* updated tutorial notebooks
* add benchmarking test coverage
* bugfixes to benchmarking framework
* ! API change: AR regularization: move from `ar_sparsity` to `ar_reg`
* documentation for Global Modeling

Included in this release:
* Now supporting use of multiple time-series datasets to train a single model (global modelling)
* example notebook for global modelling
* data frequency argument optional - automatic detection
* improved documentation and docstrings
* support for local/individual normalization of time-series when working with multiple datasets
* introducing forgetfulness: Skew model fit towards more recent observations
* widen range of default number of epochs
* add notebook for use of live-plot-loss
* improve docstrings to show up in sphinx (API documentation)
* bugfixes

* Add benchmark framework
* Support panel datasets with global modelling
* Add minimal verbosity option to fit method
* Allow no metrics
* Repeat learning-rate range test 3 times, use log10 avg
* Update energy example notebook
* Require passing dataframe for validation data while training
* Update how to build documentation added to Contributing
* Documentation using sphinx (before: mkdocs)
* Now optional: using make_future_dataframe
* avoid double calls to normalization and fill missing data methods
* New notebook guiding how to collect predictions
* Make raw predictions available to user
* Embed Tutorials in documentation page
* Embed Docstrings in documentation page
* move data to ourownstory/neuralprophet-data repository
* New energy notebook on ERCOT data
* Support more types of custom loss functions
* remove reliance on attrdict, use dataclasses instead
* improved plotting legend
* fix issues

* Robustify automatic batch_size and epochs selection
* Robustify automatic learning_rate selection based on lr-range-test
* Improve train optimizer and scheduler
* soft-start regularization in last third of training
* Improve reqularization function for all components
* allow custom optimizer and loss_func
* support python 3.6.9 for colab
* Crossvalidation utility
* Chinese documentation
* support callable loss
* Robustify changepoints data format
* require log_level in logger util
* Rename tqdm, remove overbleed option
* Reg schedule: increasing regularization in last third of training
* bug fix in plot country holidays
* Add Energy datasets and example notebook
* disable log file by default
* add double crossvalidation
* improve tests
* Buxfixes

* example notebooks: Sub-daily data, Autoregresseion
* bugfixes: `lambda_delay`, `train_speed`

* Auto-set `batch_size` and `epochs`
* add `train_speed` setting
* add `set_random_seed` util
* continued removal of `AttrDict` uses
* bugfix to index issue in `make_future_dataframe`