optimize NLL constraint propagation a little
Removes a bone-headed hot spot in NLL constant propagation; we were re-allocating the stack vector and hashmap as we repeated the DFS. This change shares those resources across each call.
It also modifies the constraint list to be a linked list; arguably I should revert that, though, as this didn't turn out to be a perf hit and perhaps the old code was clearer? (Still, the new style appeals to me.)
r? @pnkfelix
3c5f850958