Caching feature and message fingerprints

[twerkmeister]

Proposed changes:

• Caching fingerprints of messages until message is manipulated
• Caching fingerprints of features until they are manipulated

In my test on the oos dataset (roughly 15k nlu examples) the changes shaved off around 25 seconds of around 40 seconds in total for the fingerprinting. So roughly a 2/3 reduction in pure fingerprinting time.

I also looked at more general options to speed up caching, but our caching methods are just very generic and work on a very diverse set of inputs, especially arbitrarily nested dicts and lists. I found that my attempts to improve caching there, for example trying to employ frozendicts to cache some of the dictionaries, lead to longer fingerprinting times.

I also considered adding additional safeguards such as a test that checks whether all the classes methods were wrapped with either invalidates_cached_fingerprint or does_not_invalidate_cached_fingerprint. This would prevent developers from accidentally adding a new method that changes Message or Feature state and not invalidate the cached fingerprint. However, it seemed somewhat heavy handed given that it is also not a perfect protection. So I dropped it again.

Status (please check what you already did):

• added some tests for the functionality
• updated the documentation
• updated the changelog (please check changelog for instructions)
• reformat files using black (please check Readme for instructions)

[tmbo]

lgtm

[tmbo]

Nice, can we quantify this for the release briefing cc @chandrikas

[github-actions]

🚀 A preview of the docs have been deployed at the following URL: https://11596--rasahq-docs-rasa-v2.netlify.app/docs/rasa