Introduce temporary named expressions for match subjects

[sterliakov]

Fixes #18440. Fixes #17230.

#16503 did this specifically for CallExpr, but that isn't the only kind of such statements. I propose to expand this for more general expressions and believe that a blacklist is more reasonable here: we do not want to introduce a temporary name only for certain expressions that either are already named or can be used to infer contained variables (inline tuple/list/dict/set literals).

Writing logic to generate a name for every other kind of expression would be quite cumbersome - I circumvent this by using a simple counter to generate unique names on demand.

[sterliakov]

Whoa, I expected this to produce at least a couple of unused ignore diagnostics, but apparently pattern matching isn't very popular in wild...

[sterliakov]

This interferes with recently added narrowing by item/attribute (aka discriminated union, testMatchNarrowingUnionClassViaAttribute). So a simple solution like this won't fly. This information for index/attr should be stored in some other way to support narrowing of attr/item and the object itself at the same time. This may still work for other, less informative nodes (await?)

[sterliakov]

To support narrowing the subject and its "dependencies" simultaneously, I propose to just push the narrowed type of subject to typemap as the type of its original form, allowing for any inference on it.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅