Feat/onnx support

[madtoinou]

Checklist before merging this PR:

• Mentioned all issues that this PR fixes or addresses.
• Summarized the updates of this PR under Summary.
• Added an entry under Unreleased in the Changelog.

Fixes #553, fixes #2617.

Summary

• Small refactoring of the _process_input_batch and _get_batch_prediction() methods to make the behavior uniform across PLModules
• Added to_onnx method to the TorchForecastingModel class, taking care of creating the input_names and input_sample based on the model attributes
• Added a section in the User Guide to show how to export and load the model to generate forecasts

Other Information

The consistency of the forecasts was tested in a notebook for all the possible configurations, I did not implement unit-test as onnx should remain an optional dependency. Let me know what you think.

[codecov]

Codecov Report

Attention: Patch coverage is 39.39394% with 20 lines in your changes missing coverage. Please review.

Project coverage is 94.06%. Comparing base (c48521c) to head (c964e7a).

Files with missing lines	Patch %	Lines
...arts/models/forecasting/torch_forecasting_model.py	5.55%	17 Missing ⚠️
darts/models/forecasting/pl_forecasting_module.py	75.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2620      +/-   ##
==========================================
- Coverage   94.23%   94.06%   -0.18%     
==========================================
  Files         141      141              
  Lines       15509    15534      +25     
==========================================
- Hits        14615    14612       -3     
- Misses        894      922      +28     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dennisbader]

Very cool feature, thanks a lot @madtoinou 🚀 Looks like we're close already to merge :)

I had a couple of suggestions regarding the example and the export logic, where I see potential to simplify things further.

[dennisbader]

reading this I think again about how nice it would be if all our Datasets returned feature / target arrays at fixed positions (e.g. all return a tuple of past target, past cov, historic future cov, future cov, static cov, ... even if they do not support all covariate types) :D

[madtoinou]

Yeah, definitely, it would make a lot of things more intuitive and easier to tweak for users.