Fix is_a? with a union alias resolving to common ancestor

[stakach]

Resolves #9665.

Issue

A union used inline keeps its members, but the same union behind an alias was resolved to the members' common ancestor, so an unrelated sibling type matched. This affected every place a union flows through type resolution:

class Foo; end
class Bar < Foo; end
class Baz < Foo; end
class Bang < Foo; end

alias BarBaz = Bar | Baz

Bang.new.is_a?(BarBaz)    # => true  (wrong)
Bang.new.is_a?(Bar | Baz) # => false (correct)

case Bang.new
when BarBaz then ...   # matched (wrong)
end

def foo(x : BarBaz); end
foo(Bang.new)             # accepted (wrong)

begin
  raise BangError.new
rescue ex : BarBaz        # caught (wrong)
end

Fix

Aliases of unions now preserve their members, matching the inline union form, while still generalizing to the common ancestor in value contexts (pp BarBaz # => Foo is unchanged):

• is_a? and case ... when / case ... in against a union alias no longer generalize. A when Foo | Bar now matches like when Foo, Bar; unions of constants (enum members, numeric constants) keep their existing value comparison.
• Method restrictions (def foo(x : BarBaz)) restrict against the union's members.
• rescue ex : BarBaz catches the union's members rather than their common ancestor.

Covers A | B, Union(A, B), and nested-alias forms.

[ysbaddaden]

Did you consider #9665 (comment)

[stakach]

Did you consider #9665 (comment)

Yes, however the AST-node-restriction unification would be a much bigger change.

[stakach]

This is ready for review now :)