offline trace does not accurately sequence predicated memrefs

Split from #1729 as something we will not bother to solve in the near future.

With multiple predicated memrefs in a bb, the offline trace's compressed single per-bb PC entry won't identify which memref corresponds to which load or store. Thus the ifetch and memref entries inside the bb may be in a different order.

To fully handle this we'd probably want to insert a PC entry before each predicated memref. We'd put this under an option for those who care about such accuracy and are willing to pay the extra overhead.