Centralize on one payload and operation handoff between the graph and the device compilers #3
Description
The hip and cuda device are aligned, but the host device still uses a prior payload structure. Ultimately, this should all go through the DeviceOp interface instead.