Move AMDGPUAttributor run earlier in pass pipeline

The run of AMDGPUAttributor is run quite late. There isn't another run of instcombine or most other simplifications afterwards. It makes most sense to run with other IPO passes, where the full attributor would run (or, just as part of the attributor run).

If this were moved earlier, we could enable nofpclass propagation without depending on enabling the full attributor or attributor-light by default

Start by tracing where AMDGPUAttributor is scheduled in the pass pipeline and compare that position with the IPO and full attributor runs. Check the ordering of instcombine and other simplification passes, then verify that moving AMDGPUAttributor earlier enables nofpclass propagation without requiring the full attributor or attributor-light by default.

Written by the indexing model from the issue text.