perf: Iterate dataflow blocks in traversal order [pathological loop/await CFGs -52%/-82%; primary ~neutral]

[xmakro]

The dataflow fixpoint seeds its work queue in reverse postorder, but the queue itself is a plain FIFO, so on loop-heavy control flow graphs it degenerates into breadth-first iteration and revisits blocks many more times than it needs to. This replaces it with a queue that yields dirty blocks in traversal-order sweeps, tracking dirtiness in a bitset keyed by traversal position and walking it with a cursor instead of pushing and popping a deque. Blocks with no traversal position, which only backward analyses reach via predecessors of unreachable blocks, are drained separately after the ordered sweep.

A strict "always take the lowest position" rule was tried first and rejected. It is considerably worse than the current FIFO on a block joined from many back edges, such as a coroutine resume dispatch, because it returns to the join once per back edge rather than once per sweep.

The fixpoint result is unchanged: a monotone dataflow fixpoint does not depend on iteration order, and this was checked against the old queue over randomly generated control flow graphs in both directions. tests/ui and tests/incremental both pass.