Support multi-statement inference for custom types

[vnmabus]

Feature

Multi-statement inference as proposed in #254 and #1055 is supported for empty standard collection classes, like dict or list, but not for custom classes. Thus, an example such as https://mypy-play.net/?mypy=latest&python=3.10&gist=7df00fc3d8612804369dbad6b5cb973b fails.

My proposal is to allow user-defined generics to infer the generic from their first usage, so that the example work.

Pitch

This problem arose in the context of making a typed library compatible with scikit-learn (which currently is not typed). In scikit-learn, the actual data of a machine learning problem is not passed to the constructor of the object, but instead passed to the fit method (for training). Thus, if one wants to make a scikit-learn compatible object which is generic over the data accepted, one currently needs to provide the annotation, as Mypy won't infer the generic type from the call to fit. This is cumbersome for users, as most of them should benefit from the typing without altering the code to add superfluous annotations. It also obscures the examples, as adding annotations would confuse users not used to them.

[JelleZijlstra]

This would be useful, but may be difficult to implement. Mostly commenting so I can mention that mypy internally calls this feature "partial types".