Enable generic TypedDicts

[ilevkivskyi]

Fixes#3863

This builds on top of some infra I added for recursive types (Ref #13297). Implementation is quite straightforward. The only non-trivial thing is that when extending/merging TypedDicts, the item types need to me mapped to supertype during semantic analysis. This means we can't call is_subtype() etc., and can in theory get types like Union[int, int]. But OTOH this equally applies to type aliases, and doesn't seem to cause problems.

Btw the diff in semanal_typeddict.py is not as big as it looks. There was a giant if possible: ... that I found a bit annoying and replaced it with if not possible: return (IIRC GitHub has an option to ignore whitespace).

An important note on runtime support for this: IIUC this should work in Python 3.11, and should work soon with typing_extensions on earlier versions. I didn't add any version checks here. If someone is interested in adding those, please do this.

[ilevkivskyi]

It is great to have mypy_primer, will push a fix soon!

[ilevkivskyi]

The --disallow-any-generics is not strictly needed for this PR, but I found one of my new tests failed because this flag was a bit inconsistent for type aliases, so I decided to clean it up.

[ilevkivskyi]

Hm, something came to my mind, I only support class syntax for generic TypedDicts, but it seems to me we can also allow the assignment-based syntax (it would be similar to how we define generic type aliases):

TD=TypedDict("TD",{"key": int, "value": T}) # T must be unbound here.

This should be not hard to add here.

[ilevkivskyi]

Some potential merge conflicts already start to appear, it would be great if someone can take a look at this soon.

[JelleZijlstra]

Looks good, just one suggestion in the tests.

[JelleZijlstra]

Isn't this just the keys? Maybe test .values() too.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉