Unreachability should work for constrained type vars

[A5rocks]

Feature

Unreachability should be checked for constrained type vars.

Pitch

It would be nice if this raised an error when run with --warn-unreachable:

from typing import TypeVar

U = TypeVar('U', int, str)

def f(u: U) -> U:
    if u is None:
        print("whoa!!")
    
    return u

To implement this feature, I think we need to store the spans of unreachability for each run, then find the intersection. I thought for a while that we could just treat constrained type variables as if they were bound to the union of constraints, but things like #9424 will happen for things that aren't final (unlike None).

[tyralla]

I investigated a little, and this limitation seems to have been known for a while:

mypy/test-data/unit/check-unreachable-code.test

Lines 1036 to 1049 in 8241059

	def test2(x: T2) -> T2:
	if isinstance(x, int):
	reveal_type(x) # N: Revealed type is "builtins.int"
	else:
	reveal_type(x) # N: Revealed type is "builtins.str"
	if False:
	# This is unreachable, but we don't report an error, unfortunately.
	# The presence of the TypeVar with values unfortunately currently shuts
	# down type-checking for this entire function.
	# TODO: Find a way of removing this limitation
	reveal_type(x)
	return x

mypy/mypy/checker.py

Lines 1467 to 1475 in 8241059

	# We also suppress reachability warnings when we use TypeVars with value
	# restrictions: we only want to report a warning if a certain statement is
	# marked as being suppressed in *all* of the expansions, but we currently
	# have no good way of doing this.
	#
	# TODO: Find a way of working around this limitation
	if _is_empty_generator_function(item) or len(expanded) >= 2:
	self.binder.suppress_unreachable_warnings()
	self.accept(item.body)

I experimented with the IterationErrorWatcher proposed in #19270 and was successful in the first simple attempt. However, as many existing tests would require adjustment and some new ones should be introduced (which might reveal the need to adjust the method eventually), I will wait to work on it until #19270 is accepted.

By the way, aren't classes with type variables with value constraints a problem, too?