[AMD] More StreamPipeline refactoring (#7053)

The primary thing this NFC refactor does is combine 

`SmallVector<std::tuple<Operation *, int, Operation *>>
loadOpToIndLevelAndUse`

and

`MapVector<Operation *, LoadInfo> loadToInfo`

which were effectively capturing/representing the exact same thing. We
combine them by computing both the backward slice and the "info" in the
same traversal (thereby deleting `assignMemoryLayouts`, which wasn't
"assigning" anything but actually getting the `sharedEncoding` attr for
a use). Note, the boolean field `usedByDot` was not used for anything...

We also pull out this preprocessing step from `scheduleLoads` because it
doesn't actually "schedule" loads.

Author: makslevental