This is a first follow-up for#15287 (I like how my PR titles sound like research paper titles, LOL)

This PR completes the new type inference foundations by switching to a complete and well founded algorithm [1] for transitive closure (that replaces more ad hoc initial algorithm that covered 80% of cases and was good for experimenting with new inference scheme). In particular the algorithm in this PR covers two important edge cases (see tests). Some comments:

• I don't intend to switch the default for--new-type-inference, I just want to see the effect of the switch onmypy_primer, I will switch back to false before merging
• This flag is still not ready to be publicly announced, I am going to make another 2-3 PRs from the list inFoundations for non-linear solver and polymorphic application #15287 before making this public.
• I am not adding yet the unit tests as discussed in previous PR. This PR is already quite big, and the next one (support for upper bounds and values) should be much smaller. I am going to add unit tests only fortransitive_closure() which is the core of new logic.
• While working on this I fixed couple bugs exposed inTypeVarTuple support: one is rare technical corner case, another one is serious, template and actual where swapped during constraint inference, effectively causing outer/return context to be completely ignored for instances.
• It is better to review the PR with "ignore whitespace" option turned on (there is big chunk in solve.py that is just change of indentation).
• There is one questionable design choice I am making in this PR, I am addingextra_tvars as an attribute ofConstraint class, while it logically should not be attributed to any individual constraint, but rather to the full list of constrains. However, doing this properly would require changing the return type ofinfer_constrains() and all related functions, which would be a really big refactoring.

[1] Definition 7.1 inhttps://inria.hal.science/inria-00073205/document