Don't consider tyAnd/tyNot/tyOr/tyAnything as generic (#8700)

* Don't consider tyAnd/tyNot/tyOr/tyAnything as generic

`containsGenericType` was too shallow and didn't check all the branches.
The resulting half-processed nodes are often simplified by the constant
folding pass but when that's not possible we get a nasty error during
codegen.

Fixes #8693

* Move the blame onto the semFold pass

Slightly better evaluation of `is` forms.

tests/magics/t8693.nim

@@ -0,0 +1,29 @@
+discard """
+  output: '''true
+false
+true
+false
+false
+true
+true
+false
+true
+true
+'''
+"""
+
+type Foo = int | float
+
+proc bar(t1, t2: typedesc): bool =
+  echo (t1 is t2)
+  (t2 is t1)
+
+proc bar[T](x: T, t2: typedesc): bool =
+  echo (T is t2)
+  (t2 is T)
+
+echo bar(int, Foo)
+echo bar(4, Foo)
+echo bar(any, int)
+echo bar(int, any)
+echo bar(Foo, Foo)