Fix bug in lazy pattern matching

In some situations (i.e. with specialization), flambda2 might propagate
too much information. In the situation of a branch where a lazy (whose
definition by a make_block is seen by flambda2) is shortcut, and then
directly used as an integer (because the lazy wraps an enum/variant),
flambda2 might consider the branch as dead code, since the result of a
make_block (the lazy value) is used as an integer. This is obviously not
correct since the lazy value can be shortcut. The fix is to add some
opaque so that flambda2 does not propagate the information from the
make_block to the branch.

Author: Gbury

lambda/matching.ml

@@ -2130,7 +2130,7 @@ let inline_lazy_force_cond arg pos loc =
     ( Strict,
       Lambda.layout_lazy,
       idarg,
-      arg,
+      Lprim (Popaque Lambda.layout_lazy, [arg], loc),
       Llet
         ( Alias,
           Lambda.layout_int,